Separates mask and bias memory operations in attention kernel

Refactors combined mask-bias memory operations into separate dedicated operations to improve performance and maintainability.

Introduces specialized copy functions for mask and bias operations with proper bounds checking and OR-reduction for mask activity detection.

Removes redundant synchronization points by leveraging built-in synchronization in the new copy functions.

Adds predicate tensor allocation for proper boundary handling in both regular and split-KV attention kernels.

Author: algo-home