[GPU] Fix PA micro-SDPA prefill KQ OOB read

[ofirzaf]

Details:

• Fix GPU PagedAttention micro-SDPA prefill crash in sdpa_micro__prefill when the KQ microkernel handles a partial final K tile.
• The PA prefill path passes raw, unpadded KEY input, but the selected KQ GEMM strategy treated matrix A/K as padded and used block loads. This could issue out-of-bounds K reads and surface as CL_OUT_OF_RESOURCES.
• Scope the fix to PA prefill KQ only by forcing bounds-safe K access: clear A.padded, use AccessType::Scattered, and keep double masking enabled.
• Add a focused GPU unit regression:
  paged_attention_micro_sdpa_prefill_oob_sequence_test.partial_kv_tile_unpadded_key_input

Validation:

• Without the fix, the regression test reproduces the prefill failure/non-zero abort; traced during investigation to CL_OUT_OF_RESOURCES in sdpa_micro__prefill.
• With the fix, the test passes and compares against PagedAttentionReference.

Tickets:

• TBD

AI Assistance:

• AI assistance used: yes
• AI was used for root-cause investigation support and PR text drafting. The fix and regression test were manually reviewed and validated with the targeted GPU unit test.

[e-ddykim]

build_jenkins