fix: correct bidirectional attention masking in LlamaBidirectionalModel

[oliverholworthy]

Update the biencoder LlamaBidirectionalModel to handle masking correctly with newer transformers versions and different attention implementations.

• Replace the _update_causal_mask override with a proper _create_bidirectional_mask method that produces the correct mask format for each attention backend (SDPA/eager, Flash Attention 2, and the native transformers >= 5.0 path)
• The previous implementation only checked for zeros in an already-expanded mask and could silently pass incorrect mask shapes/dtypes to the attention layer

Changelog

The old _update_causal_mask had two problems:

1. Removed upstream: transformers 4.53 deleted _update_causal_mask from LlamaModel, so our override becomes a dead method that is never called. The model silently falls back to causal masking when using sdpa or eager attn_implementation
2. Wrong mask format for SDPA/eager: it returned None or passed through the raw 2D mask without expanding to 4D or casting to float, which could cause shape mismatches or dtype errors in scaled_dot_product_attention.
3. No awareness of attention backend: Flash Attention 2 expects a 2D (batch, seq_len) mask, while SDPA/eager need a 4D (batch, 1, seq_len, seq_len) float mask. The old code didn't distinguish between them.

The new _create_bidirectional_mask method:

• Uses create_bidirectional_mask from transformers.masking_utils when available (transformers >= 5.0), which handles all backends natively.
• Falls back to backend-aware logic for older transformers: passes the 2D mask directly for flash_attention_2, and uses _prepare_4d_attention_mask for SDPA/eager.

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?

If you haven't finished some of the above items you can still open "Draft" PR.

Additional Information

• Related to # (issue)