llama.cpp adds flash-attn for Mistral Small 4

This is the kind of low-level kernel work that quietly changes real-world model performance: one missing attention shape can erase most of the benefit of GPU offload.

• –The patch adds MMA-f16 and tile kernel configs, dispatch logic, template instances, and a dedicated tile .cu path for the 320/256 shape.
• –The key constraint is ncols2=32, which matches the model’s GQA ratio and narrows the fix to the exact path Mistral Small 4 needs.
• –Benchmarks in the PR show a dramatic gap versus fallback behavior: prefill jumps from roughly 180 tok/s to about 3.7k tok/s, and decode from about 33 tok/s to about 186 tok/s.
• –The upside is obvious for CUDA users running this model; the downside is that the fix is highly shape-specific, so broader coverage will still matter for future variants.