[GPU] Support IncreasePositionIdsPrecisionForGemma4 for PA backend

[ahnyoung-paul]

Details

Problem: The IncreasePositionIdsPrecisionForGemma4 pass introduced in PR #36896 does not match the rotary embedding graph pattern when using the PA (Paged Attention) backend. Under PA, upstream transformations fold Transpose into Reshape and Unsqueeze into Reshape, producing a different subgraph shape (MatMul → Reshape → Concat → Sin/Cos → Reshape → RoPE) that the original pattern (MatMul → Convert → Transpose → Concat → Sin/Cos → Unsqueeze → RoPE) cannot match. As a result, the rotary embedding MatMul remains in FP16 and the infinite repetition loop persists on PA backend.

Fix: Updated the IncreasePositionIdsPrecisionForGemma4 MatcherPass pattern to accept the PA backend's graph structure:

• Added MatMul → Reshape as an alternative concat input path (alongside the existing MatMul → Convert → Transpose)
• Added Sin/Cos → Reshape as an alternative RoPE input path (alongside the existing Sin/Cos → Unsqueeze)
• Made the Convert after MatMul optional in the callback (present in SDPA path, absent in PA path)

Tickets

• CVS-187785

AI Assistance

• AI assistance used: yes
• AI was used for per-iteration logit comparison analysis, tensor dump comparison scripting, exec graph subgraph extraction, and matcher log interpretation. All code changes were manually reviewed and validated with end-to-end accuracy tests.