Adds varlen FDMA and padding-free attention

Introduces lazy resolution of attention kernels and padding helpers, plus a compile-friendly kwarg processor that adapts to kernel feature support.

Enables variable-length execution via unpad/repad when masks are 2D, and padding-free/packed flows using position ids or precomputed sequence offsets. Adjusts is_causal for single-token queries and supports windowed attention with bias-safe top-k selection.

Improves compatibility across kernel versions and torch.compile, adds deterministic control via env var, handles PEFT dtype quirks, and includes minor device safeguards. Raises a clear error when incompatible mask/bias shapes are mixed.

Author: algo-home