[CPU] Fix SDPA node attention mask precision handling for bf16/f16 inference

[liubo-intel]

Details:

• Use actual attention mask input precision instead of compute precision (bf16/f16) to fix LFM2-350M output corruption when running with low precision on Xeon platforms.

Tickets:

• CVS-177340

[Copilot]

Pull request overview

This PR fixes precision handling for attention masks in the Scaled Dot Product Attention (SDPA) implementation when running with BF16/F16 inference precision on CPU. The issue caused output corruption in LFM2-350M model on Xeon platforms.

Key Changes:

• Uses actual attention mask input precision instead of assuming compute precision (bf16/f16)
• Fixes pointer arithmetic to use byte-based strides for attention masks
• Adds comprehensive test coverage for stateful SDPA with boolean masks in BF16 precision

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
stateful_sdpa_bool_mask.cpp	Adds new test case validating SDPA with boolean masks in BF16 inference mode
scaled_attn.cpp	Fixes attention mask precision detection and pointer arithmetic for both ONEDNN and ACL kernel implementations

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[zhangYiIntel]

LGTM

[zhangYiIntel]

Could we reuse existing tests to cover the boolean attn_mask ? The only change here is the attn_mask type.

[zhangYiIntel]

Please also add copy right in the head of the file.

[liubo-intel]

Could we reuse existing tests to cover the boolean attn_mask ? The only change here is the attn_mask type.

previously I try to extend current existing SDPA tests to cover our changed logic, but I found these SDPA tests are stateless and will be converted to Snippets(Subgraph) during inference which will not use our changed logic, so I created new ones to cover our changes.