[None][perf] triton paged attention: non-pow2 head_dim, decode speedup, logit cap, FP8 KV cache support

- HEAD_DIM_PADDED: pad non-power-of-2 head dims to next power of 2 for Triton efficiency
- WRITE_DIRECT optimization: skip intermediate buffer for direct output writes
- KV offset hoisting: precompute KV offsets outside inner loop
- SW-aware splits: sliding window-aware chunking for context attention
- Logit cap support: per-head logit softcapping (Gemma4 alt-attention)
- Always-Triton threshold: route to Triton for all seq_lens above threshold
- Two-chunk gather: efficient gather for non-pow2 head dims in stage2
- FP8 KV cache casting: cast K/V to query dtype at all load sites (decode
  SW path, context block_ptr loads, stage2, fallback SDPA path)

Signed-off-by: Chenghao Zhang <211069071+nvchenghaoz@users.noreply.github.com>

Author: nvchenghaoz