[CUDA] Reduce the number of stream-k blocks to reduce the overhead of the flash_attn_stream_k_fixup kernel#21086

gaugarg-nv wants to merge 3 commits intoggml-org:masterggml-org/llama.cpp:masterfrom

gaugarg-nv:reduce_stream_k_blockgaugarg-nv/llama.cpp:reduce_stream_k_blockCopy head branch name to clipboard