Adds Triton dynamic mask attn kernels

Enables fused Triton forward/backward paths for dynamic masked attention to reduce padding overhead and deliver faster windowed attention execution.

Author: LoserCheems