Add varlen MHA fwd/bwd; enable paged KV (no mask)

Re-enables variable-length attention forward/backward and registers both with the extension.
Simplifies the varlen API by removing mask/bias; uses empty placeholders and flags, and drops dbias from outputs.
Enables paged KV cache for varlen forward, validates left padding, preserves zero_tensors/deterministic handling, and applies minor formatting cleanups.

Author: algo-home