fix(retrieval): export deployable Sentence Transformers checkpoints

[oliverholworthy]

What does this PR do ?

Make Ministral retrieval fine-tuning produce a stock Hugging Face text checkpoint with generated Sentence Transformers metadata that reflects the trained encoder behavior.

Changelog

• Load standard ministral3 embedding backbones with the stock Hugging Face class and is_causal=false; retain the custom ministral3_bidirec class only for explicit and legacy checkpoints.
• Generate Sentence Transformers modules, pooling, normalization, prompts, similarity, and sequence-length metadata from the effective bi-encoder at save time.
• Restrict standard Sentence Transformers export to representable text encoders and preserve the existing save path for unsupported pooling and multimodal backbones.
• Reuse the resolved local source revision and tokenizer assets without a second model download, while removing stale multimodal processor metadata from the extracted text checkpoint.
• Preserve size-based consolidated sharding when an extracted text model no longer matches its parent VLM shard map.
• Use the native Mistral tokenizer in the Ministral retrieval recipe and accept the collator no-op return_token_type_ids=false argument.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation? (No separate documentation changes are required; the example recipe is updated.)

If you have not finished some of the above items you can still open "Draft" PR.

Additional Information

• uv run ruff format --check . passes.
• uv run ruff check . passes.
• The focused retrieval suite passes 18 tests, including a CPU regression that exercises the real Transformers FlashAttention adapter and verifies the kernel receives is_causal=false.
• The consolidated shard-index and size-allocation suite passes 10 tests.
• A one-step Ministral 3B bi-encoder fine-tune completed on Transformers 5.12.1 using SDPA and produced a consolidated checkpoint.
• The checkpoint loads locally with stock Ministral3Config / Ministral3Model, is_causal=false, no custom auto_map, and Sentence Transformers 5.4.0 successfully encodes a query to a 3072-dimensional embedding.
• Source tokenizer and chat-template assets are retained; stale Pixtral processor metadata is intentionally omitted because the exported checkpoint is text-only.
• No linked issue.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[oliverholworthy]

/ok to test a9f4423