llama.cpp adds flash-attn support

A flash-attn kernel for DKQ=320 / DV=256 has been added to ggml-cuda. The key point is limiting to ncols2=32, moving the CPU fallback that occurred with Mistral Small 4's GQA=32 combination onto the CUDA path.

The patch includes MMA-f16 and tile kernel config, dispatch logic, template instances, and a tile .cu file for Mistral Small 4.

Benchmarks were measured on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition.

• Before the change: pp512 179.64±2.66, tg32 33.05±0.30
• ncols2=32: pp512 3752.47±26.33, tg32 185.92±4.43
• ncols2=8: pp512 3678.35±21.68, tg32 182.24±4.00

In a follow-up fix, a bug where the per-warp-group base index was missing at ncols=32, causing results to diverge from the CPU output, was fixed, and a FLASH_ATTN_EXT failure involving sinks=1 was also resolved. In the latest retest, pp512 3655.55±15.81 and tg32 186.31±4.23 were confirmed.

In multi-GPU environments as well, pp512 improved from 348.35 → 1310.13 and tg128 from 59.12 → 93.06.