coleased memory reads for faster backward pass in attention

[ngc92]

Uses one warp (instead of one thread) for each result that is to be computed. We gain coalesced access in the inner loop, translating to a tremendous speedup.