Unifies attention kernels with bias+mask windowing

Refactors attention paths to accept external attention bias and boolean causal mask, replacing zoh/dt-based masking and cache-position logic. Introduces a generic mask preparer that applies top-k windowing (optionally causal-aware), and standardizes interfaces across SDPA, Flash, Triton, and Flex implementations.

Removes zoh/dt projection and related params, repeats KV artifacts for GQA, and consistently applies additive masks. Updates benchmarks to generate bias/mask inputs, rename keep_window_size to window_size, adjust head dims, and harmonize result handling and output labeling.

Improves API consistency, simplifies experimentation with custom biases, and aligns masking semantics across kernels for more reliable benchmarking.

Author: algo-home