refactor: unify rec multi round decode mode with one-stage flag.

[LMX-xin]

What This PR Changes

1. Uses a single gflag to select decode mode:
   • FLAGS_enable_xattention_one_stage == true -> one-stage
   • FLAGS_enable_xattention_one_stage == false -> two-stage
2. Removes legacy/extra mode-selection paths:
   • enable_xattention_two_stage_decode
   • enable_xattention_one_stage_decode
   • is_xattention_two_stage_decode_enabled()
3. Removes cache-state-based mode gating:
   • No branch selection based on xattention_two_stage_decode_cache.has_value()
   • No branch selection based on aggregated two_stage_* .defined() checks

[LMX-xin]

I will rebase after #933 merge

[gemini-code-assist]

Code Review

This pull request introduces a significant refactoring to unify the multi-round decode modes for rec under a single flag, FLAGS_enable_xattention_one_stage. The changes are extensive, touching attention kernels, metadata builders, and the CUDA graph executor. The two modes, one-stage and two-stage decode, are now cleanly separated, with the two-stage path implementing a shared/unshared attention optimization. New components like XAttentionWorkspace and xattention_planinfo have been added to support this. Crucially, a new test has been added to compare the outputs of both decode paths, ensuring correctness of the refactoring. The implementation appears solid and consistent across the codebase. I have no high or critical severity comments.

[DragonFive]

LGTM