Support Qwen3.5 embedding LoRA adapters with vLLM

## Summary

Add end-to-end `train_unembed` support for Qwen3.5-based models (including Qwen3.6 and Qwen3.8, which use the same Transformers architecture) by training and serving a LoRA adapter on `embed_tokens`, matching Tinker's public API behavior.

This is intentionally separate from #149 / #150, which add the Gated DeltaNet attention targets. Until embedding-adapter serving is available, TuFT accepts the SDK-default `train_unembed=True` for Qwen models but does not add an embedding target.

## Upstream dependency

TuFT currently pins vLLM 0.24.0. vLLM has generic embedding-LoRA support, but the Qwen3.5 model did not register its embedding modules in v0.24.0 or v0.26.0.

The required Qwen3.5 registration was merged upstream in vLLM PR [#48850](https://github.com/vllm-project/vllm/pull/48850). Prefer waiting for and upgrading to a vLLM release that contains that change, rather than depending on an unreleased commit. If a backport is chosen instead, document that decision explicitly.

## Implementation notes

- Map `train_unembed=True` to `embed_tokens` for the Qwen3.5 architecture without affecting the vision encoder.
- Keep HF and FSDP target resolution and saved adapter metadata aligned.
- Extend TuFT's vLLM checkpoint-key compatibility for the flat Transformers `model.embed_tokens` path versus vLLM's nested `language_model.embed_tokens` path.
- Verify tied-embedding/base-weight behavior in PEFT checkpoints.
- Add an end-to-end GPU test that trains, exports, loads, and serves the embedding adapter through vLLM.
- Update the LoRA target-module documentation once support is enabled.

## Acceptance criteria

- A Qwen3.5-based run with `train_unembed=True` actually trains `embed_tokens`.
- The resulting adapter loads and serves successfully in the supported released vLLM version.
- HF and FSDP resolve the same embedding target and neither path targets the vision encoder.
- Existing Qwen attention/MLP and Gated DeltaNet LoRA behavior remains unchanged.