feat(detection): route vision-stripped VLM checkpoints to a text-only path

## Summary

Community checkpoints increasingly ship a VLM `model_type` with the vision tower stripped: the weights hold only `language_model.*` and the config either omits `vision_config`, sets it to `{}` / `null`, or keeps it and sets `language_model_only: true`. mlxcel handles the first shape for four families (`qwen3_5`, `qwen3_5_moe`, `gemma3`, `gemma4`) through `detect_text_or_vlm`, and fails on every other shape: an empty or null `vision_config` still counts as "has vision" and the VLM loader then errors on the sub-config, `language_model_only: true` is a deliberate named error, and the Qwen2-VL / Qwen2.5-VL / Qwen3-VL / Qwen3-VL-MoE families have no text-only route at all. This change makes detection key on the presence of vision weights and makes the Qwen VL family loaders tolerate a missing tower.

## Current behavior

`src/models/detection.rs:55-57`:

```rust
pub(crate) fn has_vision_config(config: &serde_json::Value) -> bool {
config.get("vision_config").is_some()
}
```

so `"vision_config": {}` and `"vision_config": null` route to the VLM loader, which then fails in `parse_required_vlm_subconfig` (`src/loading/vlm.rs:177-187`, `Failed to parse ... vision config`) or in the encoder constructor on the first missing `vision_tower.*` tensor (`src/loading/vlm_qwen.rs:260-272` for `qwen3_vl`, `:212-226` for `qwen2_5_vl`, `:56-75` for `qwen2_vl`, `:306-318` for `qwen3_vl_moe`).

`src/models/qwen3_5.rs:236-251` `validate_qwen35_wrapper_config` returns `Qwen35UnsupportedConfig::LanguageModelOnly` for `language_model_only: true`, even when the checkpoint is exactly the vision-stripped build the flag describes. The text loader for the same family already strips `vision_tower.*` / `model.visual.*` and accepts `language_model.*` keys (`src/models/qwen3_5.rs:3140-3158` `sanitize_moe_weights`, `src/loading/special.rs:44-58` `qwen35_text_config`), so the routing is the only obstacle for that family.

`gemma4` already decides by weights (`src/models/detection.rs:59-79` `gemma4_has_vision_weights` reads `model.safetensors.index.json`), which is the pattern to generalize.

A stripped Qwen3.6 build (`leonsarmiento/ThinkingCap-Qwen3.6-27B-3bit-mlx`: `model_type qwen3_5`, no `vision_config`, `language_model_only: false`, weights under `language_model.*` only) loads today because it happens to omit `vision_config`; the same weights with `vision_config: {}` or `language_model_only: true` do not.

## Expected behavior

Detection decides "text-only" for a VLM `model_type` when any of the following holds:

1. `vision_config` is absent, `null`, or an empty object;
2. top-level `language_model_only` is `true`;
3. the safetensors index (or the single `model.safetensors` header) contains no key starting with `vision_tower.`, `model.visual.`, `model.vision_tower.`, `visual.`, or `vision_model.`.

When text-only is decided for `qwen3_5` / `qwen3_5_moe` / `gemma3` / `gemma4`, route to the existing text `ModelType`. For `qwen2_vl`, `qwen2_5_vl`, `qwen3_vl`, `qwen3_vl_moe` (whose text decoders carry MRoPE and cannot reuse `Qwen2` / `Qwen3`), keep the VLM `ModelType` but load with `vision_encoder: None`; an image or video request on such a model returns the error `model was loaded without a vision tower (text-only checkpoint)` instead of panicking, and text requests run on the text-only fast path that already exists (`can_use_text_only_fast_path`).

A config that says `language_model_only: true` but ships vision weights loads text-only (the flag wins) with a one-line stderr notice naming the flag.

## Implementation plan

1. `src/models/detection.rs`: replace `has_vision_config` with `fn vlm_has_vision(config: &Value, model_path: &Path) -> bool` implementing rules 1-3 (reuse `gemma4_has_vision_weights` as the weight scan, generalized to the prefix list above; also scan a single-file `model.safetensors` header when no index exists). Use it in `detect_text_or_vlm` (pass `model_path`) and for the Qwen VL arms to set a new `ModelType` flag or a `LoadOptions { text_only: bool }` carried to the loader.
2. `src/models/qwen3_5.rs:236-251`: `language_model_only: true` stops being an error on the VLM loader because detection no longer sends such configs there; keep the wrong-type error.
3. `src/loading/vlm_qwen.rs`: in `load_qwen2_vl`, `load_qwen2_5_vl`, `load_qwen3_vl`, `load_qwen3_vl_moe`, when `text_only`, skip `parse_required_vlm_subconfig("vision_config")` and the encoder constructor; build the model with `vision_encoder: None` and a processor built from defaults for `spatial_merge_size` (read from `vision_config` when present, else 2).
4. `src/vision/qwen2_vl.rs:36`, `src/vision/qwen2_5_vl.rs:36`, `src/vision/qwen3_vl.rs:38`, `src/vision/qwen3_vl_moe.rs:32`: change `vision_encoder` to `Option<...>`; `get_input_embeddings` / `input_embeddings_multimodal` return `Err` (or the runtime checks `has_vision_tower()` before calling) with the message above. Add `fn has_vision_tower(&self) -> bool` to `VlmRuntimeRef` so `src/multimodal/vlm_runtime.rs` and `src/server/model_worker.rs` reject media requests cleanly; `mlxcel list` / model capabilities report `vision: false` for such a load.
5. Tests in `src/models/detection_tests.rs`: `empty_vision_config_routes_to_text` (`qwen3_5` with `vision_config: {}`), `null_vision_config_routes_to_text`, `language_model_only_routes_to_text`, `no_vision_weights_in_index_routes_to_text` (temp dir with an index holding only `language_model.*` keys), `qwen3_vl_without_vision_config_loads_text_only` (loader-level, synthetic tiny weights).
6. `docs/supported-models.md`: a short paragraph under the Qwen VL and Qwen3.5 entries describing text-only loads, and remove the sentence that `language_model_only: true` is rejected.

## Validation

(a) Unit tests as in step 5.

(b) Real checkpoint: `leonsarmiento/ThinkingCap-Qwen3.6-27B-3bit-mlx` (12.5 GB, stripped Qwen3.6-27B; `model_type qwen3_5`). Edit a local copy's `config.json` to add `"vision_config": {}` and, separately, `"language_model_only": true`, and confirm all three variants load and produce identical text:

```
./target/release/mlxcel generate -m models/ThinkingCap-Qwen3.6-27B-3bit-mlx -p "Explain why the sky is blue in two sentences." -n 64
```

Acceptance: the three config variants produce byte-identical output; an `--image` request on the loaded model returns the text-only error rather than a panic; `mlxcel list` still reports the family.

## Acceptance criteria

- [ ] `vision_config` absent / null / `{}` and `language_model_only: true` all route VLM `model_type`s to a text-only load
- [ ] A missing vision tower in the weight index routes to text-only for the families above
- [ ] Qwen2-VL / Qwen2.5-VL / Qwen3-VL / Qwen3-VL-MoE load with `vision_encoder: None` and reject media requests with a named error
- [ ] Stripped Qwen3.6 checkpoint loads under all three config variants with identical output
- [ ] docs/supported-models.md updated
- [ ] detection table in src/models/detection.rs updated with a test
- [ ] cargo test --workspace --profile test-fast --features metal,accelerate passes
- [ ] cargo clippy --workspace --all-targets -- -D warnings and cargo fmt --all -- --check pass

## Out of scope

- Text-only loads for encoder-based VLM families without a reusable text decoder route (LLaVA, Idefics, Molmo, Pixtral); each needs its own follow-up once a stripped checkpoint exists.

Start with src/models/detection.rs and its detection tests, especially the existing gemma4 weight scan, then trace the Qwen VLM loaders in src/loading/vlm_qwen.rs. Review the four Qwen vision runtime files and src/multimodal/vlm_runtime.rs for the text-only and media-error paths. Done means the listed synthetic tests, real checkpoint validation, documentation update, and workspace test, clippy, and format checks pass.

Written by the indexing model from the issue text.