Update on "Limit softmax to causally-valid elements in cpu_sdpa"

Instead of setting masked positions to -inf and computing
max/exp/normalize over all kvSize elements, limit the softmax
to only the causally-valid range per row. This avoids unnecessary
computation on masked positions and zero-fills them for GEMM 2.

Differential Revision: [D96044307](https://our.internmc.facebook.com/intern/diff/D96044307/)

[ghstack-poisoned]

Author: kimishpatel