docs(embeddings): the detection table omits model_type aliases the loader accepts

## Summary

The `docs/embeddings.md` detection table omits `model_type` aliases the loader actually accepts, so a reader whose HF export spells `xlm_roberta` (underscore) or `siglip_text_model` concludes their checkpoint is unsupported when it serves fine.

## Background

The detection code accepts both the hyphen and underscore spellings for these families, but the docs list only one spelling each. The rest of the page checks out, so this is a narrow, mechanical fix.

## Proposed Solution

Add the alias spellings to the two table cells and the one sentence, verified against the `match` arms in `detection.rs`.

## Implementation Notes

- `docs/embeddings.md:83` lists only `xlm-roberta`; `src/models/detection.rs:428` accepts `"xlm-roberta" | "xlm_roberta"`.
- `docs/embeddings.md:85` lists only `siglip`; `src/models/detection.rs:430` accepts `"siglip" | "siglip_text_model"`.
- `docs/embeddings.md:76` says `ForSequenceClassification` detection covers "bert, xlm-roberta and modernbert"; `src/models/detection.rs:347-348` defines `RERANKER_CLASSIFIER_MODEL_TYPES = &["bert", "xlm-roberta", "xlm_roberta", "modernbert"]`.
- The rest of the page (the source-map paths, architecture list, and numeric constants) matches the code.

## Acceptance Criteria

- [ ] Every `model_type` spelling `detection.rs` accepts for these families appears in the table.

---

## Original Suggestion

### Title: docs(embeddings): the detection table omits model_type aliases the loader accepts

The `docs/embeddings.md` detection table omits `model_type` aliases the loader actually accepts, so a reader with an HF export that spells `xlm_roberta` (underscore) or `siglip_text_model` concludes their checkpoint is unsupported when it serves fine.

## Evidence

- `docs/embeddings.md:83` lists only `xlm-roberta`; `src/models/detection.rs:428` accepts `"xlm-roberta" | "xlm_roberta"`
- `docs/embeddings.md:85` lists only `siglip`; `src/models/detection.rs:430` accepts `"siglip" | "siglip_text_model"`
- `docs/embeddings.md:76` says `ForSequenceClassification` detection covers "`bert`, `xlm-roberta` and `modernbert`"; `src/models/detection.rs:347-348` — `RERANKER_CLASSIFIER_MODEL_TYPES = &["bert", "xlm-roberta", "xlm_roberta", "modernbert"]`

(The rest of the page checks out: the ~41 source-map paths, the architecture list, and the numeric constants at `:107,119,123` all match the code.)

## Suggested fix

Add the alias spellings to the two table cells and the one sentence — verified by reading a single `match` arm in `detection.rs`.

## Acceptance criteria

- [ ] Every `model_type` spelling `detection.rs` accepts for these families appears in the table

Start with the detection table and sentence in docs/embeddings.md, then verify the accepted spellings in the match arms and RERANKER_CLASSIFIER_MODEL_TYPES in src/models/detection.rs. Add every listed alias for xlm-roberta and siglip to the documentation, then confirm the acceptance criterion and review the surrounding table for consistency.

Written by the indexing model from the issue text.