fix: support external page tables with 4-byte alignment

[rogeryoungh]

Summary

SGLang integration requires two changes to MSA sparse attention:

• Accept a prebuilt page_table directly.
• Remove the 16-byte alignment assumption for page-table metadata and use the natural 4-byte alignment of int32.

Changes

• Added page_table support to the public API and sparse prefill adapter, including mixed-batch routing.
• Centralized page-table validation for BF16 and NVFP4 paths.
• Updated the CuTe tensor alignment and compilation cache keys.
• Added coverage for direct page tables, mixed dispatch, precedence, and non-16-byte-aligned inputs.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}