[None][perf] triton paged attention: non-pow2 head_dim, decode speedup, logit cap

Several performance improvements to the Triton paged attention kernel:

1. HEAD_DIM_PADDED support for non-power-of-2 head dimensions (e.g. Gemma 4's
   head_dim=176=128+64 and head_dim=512). The decode kernel handles the non-pow2
   head_dim by storing partial results in a HEAD_DIM_PADDED staging area and
   gathering them at stage2 combine time, eliminating zero-padding overhead.

2. Two-chunk gather kernel for non-pow2 head_dim: processes the two chunks
   in separate vectorised passes to improve memory bandwidth utilisation.

3. WRITE_DIRECT optimisation: when num_splits==1 the stage-2 combine is
   short-circuited and the single-split result is written directly to the
   output tensor, saving one kernel launch per decode step.

4. KV offset hoisting: loop-invariant K/V tensor offsets and block pointers
   are hoisted out of the inner attention loop — 9-22% decode speedup on
   H100/B200 benchmarks.

5. Sliding-window-aware split computation: the number of KV splits is capped
   by the sliding window size rather than the full sequence length, reducing
   wasted compute for SW attention models.

6. Logit soft-cap support via a logit_cap parameter for models that apply
   logit capping (e.g. Gemma 4 with logit_softcapping=30.0).

7. SLIDING_WINDOW added to autotune key; num_stages/num_warps search space
   extended for better occupancy on newer hardware.

8. Always-Triton threshold: raised _TRITON_T_THRESHOLD so Triton paged
   attention is always used regardless of batch size, avoiding extra
   FlashInfer dispatch overhead in piecewise CUDA graph workloads.

Test: updated test_triton_paged_attention.py to cover new configurations.
Signed-off-by: Chenghao Zhang <211069071+nvchenghaoz@users.noreply.github.com>

Author: nvchenghaoz