feat: support ONNX models that require token_type_ids

[j-sperling]

Summary

• BERT-family ONNX exports declare a token_type_ids input, and session.run() requires every declared input in the feed — so models like Xenova/all-MiniLM-L6-v2 and Xenova/bge-small-en-v1.5 currently fail with Required inputs (['token_type_ids']) are missing. This detects declared inputs once at init and feeds all-zero segment ids when required (correct for single-sequence embedding). XLM-R-family models (the bge-m3 default) are unaffected.
• Why it matters: it unlocks a small-model tier for bulk ingest without touching the default. Verified locally: Xenova/bge-small-en-v1.5 (33M, CLS-pooling-native — matching this provider's CLS pooling) loads, passes a semantic sanity check, and embeds a 64-text batch in ~0.02s vs ~6.6s for the 568M int8 bge-m3 default on the same CPU. Corpus-scale context: indexing 186K chunks with the default took >4.5h on an M-series CPU; a small-tier model brings that to minutes via embedding.model config, no code change.
• Deliberately not changing DEFAULT_MODELS["onnx"]: a default swap changes embedding dimension (1024 → 384) and would break existing Milvus collections on upgrade. Users opt in per install via memsearch config set embedding.model.
• Caveat worth documenting: Xenova/multilingual-e5-small now loads but is not recommended — e5 models expect query:/passage: prefixes and mean pooling, and scored poorly on the sanity check under CLS pooling.

Test plan

• tests/test_embeddings_onnx_inputs.py: stub-session tests that token_type_ids is fed as zeros (shape-matched to input_ids) when declared, and omitted when not — no model download needed
• Full suite: 238 passed, 7 skipped
• ruff check / ruff format --check clean
• Live load + embed + semantic sanity for bge-small-en-v1.5, all-MiniLM-L6-v2, multilingual-e5-small; default bge-m3 path unchanged