flash-attention

Bump to v2.5.4

Update Cutlass to v3.4.1

Don't need to reduce row_sum during online softmax

Optimize compile to 1: avoid oom 2: minimize swap usage 3: avoid threads starvation when letting ninja decide how many workers to spawn or manual MAX_JOBS "guesses". Logic is to take the min value of MAX_JOBS auto-calculated by two metrics: 1: cpu cores 2: free memory. This should allow flash-attn to compile close to the most efficient manner under any consumer/server env. (#832)

Bump to v2.5.3