UPSTREAM PR #17606: vulkan : fix FA mask load with bounds check (coopmat2)

[loci-dev]

Mirrored from ggml-org/llama.cpp#17606

Bug was caused by variable shadowing in flash_attn_cm2, introduced in #17186

It was only observable with coopmat2 and GGML_KQ_MASK_PAD=1

Repro:

1. modify ggml.h and set GGML_KQ_MASK_PAD 1
2. run test-backend-ops-o FLASH_ATTN_EXT

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #369

Analysis: Vulkan shader bug fix removing variable shadowing in Flash Attention cooperative matrix implementation.

Performance Impact: Zero measurable change across all binaries. Power consumption analysis shows 0.0% variation in all 16 binaries (libllama.so: 193,067 nJ unchanged, llama-run: 192,101 nJ unchanged). No functions exhibit response time or throughput changes.

Code Change: Single-line modification in flash_attn_cm2.comp removes redundant mv variable declaration that shadowed outer-scope variable. This correctness fix eliminates undefined behavior when GGML_KQ_MASK_PAD=1 is enabled with Vulkan coopmat2 backend.

Inference Impact: No tokens-per-second impact. Core inference functions (llama_decode, llama_encode, llama_tokenize) show identical execution profiles between versions.