Unifies dmattn to bias+mask API; expands tests

Replaces internal zoh/value-proj masking with an external attention bias plus optional causal mask and top‑k windowing, simplifying the interface and masking semantics across backends.

Aligns Python, CUDA, Triton, and Flex to a shared signature, applies masking consistently, ensures contiguous layouts, and uses deterministic execution for stable gradients.

Expands backward‑equivalence coverage to head dims 192/256 and updates tests to use bf16 bias and causal masks, improving reproducibility and backend parity.

Author: algo-home