[CPU][PA] Align bidirectional image attention and sliding window interaction with reference

[p-wysocki]

Details:

• When both bidirectional image attention and a sliding window were active, the CPU plugin computed the attention start index as:
  start_idx = image_group_end - sliding_window
  Whereas the correct way to do it (implemented by transformers and GPU Plugin) is to not clip the image attention, but instead pass it as a whole block, no matter what the sliding window size is.

• An example

Let's assume 6 image tokens and a sliding window of size 5.

Before the fix:
Attention mask: [1, 6) - first token cut

After the fix:
Attention mask: [0, 6) - full group

If attention mask is being calculated for a text token, the attention is regular causal/sliding window, no matter if previous tokens were image or text, which matches transformers implementation.

Tickets:

• CVS-185393

AI Assistance:

• AI assistance used: yes, to write tests and make sure the changes don't impact other aspects/inputs of PA

[maxnick]

@zhangYiIntel , could you please review?