Merge main into asonawane/main-ops-sparse-attention-indexer
Preserve state_status at output 8, append snapshot outputs at 9-11,
and retain BF16 prefill dispatch with an explicit quantized-cache guard.
Add coverage for snapshot outputs with and without state_status.
Validated using isolated merged-source provider/test links: 119 passed,
2 architecture-specific skips on GPU 1. Active worktrees are unchanged.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
