Add page-16 sliding-window attention support

[zjing14]

Summary

Extend FlyDSL paged attention for SBD-style short-query workloads:

• Support page sizes 16 and 64 for vectorized paged BF16/F16 attention.
• Add causal sliding-window masking and KV tile pruning.
• Route page-size-16 FP8 sliding-window decode through the tile kernel.
• Support query lengths 1–16 for FP8 paged decode.
• Package the kernels modules in wheels and source distributions.

The existing dual-wave path remains page-size-64 only. Page-size-16 and sliding-window workloads use the short-query generic path.

Testing

• Page-size-16 vectorized BF16 attention:
  • Query lengths 1–16
  • Global and sliding-window attention
• Page-size-16 FP8 paged decode:
  • Query lengths 1, 15, and 16
  • Global and sliding-window attention

python -m pytest -q \
  tests/kernels/test_flash_attn_fwd.py \
  tests/kernels/test_pa.py \
  -k 'test_vectorized_paged_page16_short_query or test_small_block_accuracy'

Result: 38 passed.

- Black and Ruff checks passed.
- git diff --check passed.
- Built and inspected the source distribution; required kernels.attention and kernels.common modules are included.
- Validated downstream with FlyDSL FP8 attention and SBD on 4× MI355X.