Fix: Qwen3 reranker tokenizer lacks pad_token for batch requests

[maxmax117]

when you use qwen3 reranker and a batch requests, it always fails for lack of pad_token, to work around you would have to submit the the batch requests one by one ,which is low efficiency. please take a look and fix it, it should be a easy fix.

[joonsoome]

@maxmax117

Thanks for the report! I can reproduce the batch failure when the Qwen3 reranker tokenizer has no pad_token and padding=True is used. (It could be a model characteristics)

Plan:

1. Add a pad_token fallback after tokenizer load (prefer eos_token, else unk_token; warn if both missing).
2. Apply the fix in both MLX reranker and Torch CrossEncoder backends.
3. Add a small test to ensure batch rerank works when pad_token is missing.
4. Update troubleshooting docs with the fallback behavior.

For sure, if you can share the exact reranker model id and the error stack trace, I can validate the fix against your setup.