Fused Softmax in Triton

Mid-week 2nd kernel: fused softmax. Naive softmax needs 3 passes over memory (max, exp+sum, divide) — fusing into one pass takes it from ~55 GB/s to ~230 GB/s, a 4x jump.

Triton's line is also tighter and more consistent than PyTorch's across block sizes.

Next: flash attention.