[FlashInfer MLA] Use public CUDA graph plan updates

[saltyminty]

Motivation

Replace SGLang's private FlashInfer MLA CUDA-graph replanning bridge with the
public plan-update API introduced by
FlashInfer #5041.

The current implementation monkey-patches plan() and reaches into private
workspace, metadata-buffer, and generated-module attributes. This draft shows
the public replacement while preserving the existing eager and full-plan paths.

This PR remains a draft until SGLang can consume a FlashInfer release containing
the new API. Validation below used FlashInfer commit
14096da3f1490d94a640c7d55a07bbe856f20279 directly.

Modifications

• Opt CUDA-graph MLA wrappers into public graph plan updates.
• Submit complete MLAPlanMetadata.csr(...) values through
  update_cuda_graph_plan() during decode and target-verify replay.
• Keep a reusable non-alias page-index source for the public update contract.
• Remove both private fast_mla_* planner helpers and their plan()
  monkey-patches.
• Leave existing eager planning and attention runtime calls unchanged to keep
  this downstream demonstration narrowly scoped.

Accuracy Tests

Targeted B200/SM100 validation against the FlashInfer commit above:

• test/registered/attention/unittests/mla/test_flashinfer.py: 6 passed,
  1 registered skip, and 15 subtests passed. The skip is the existing SM100
  EAGLE draft CUDA-graph known failure.
• Task-local public-update probe: 20/20 updates passed, capture-stable buffer
  pointers remained unchanged, the page-index source was non-aliasing, and no
  private bridge symbols remained.

Speed Tests and Profiling

No standalone SGLang speed benchmark was run for this downstream demonstration.
The sync-free implementation and performance validation live in FlashInfer
#5041; this PR only replaces SGLang's private bridge with that public API.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. (The existing registered MLA suite exercises the changed paths.)
• Update documentation according to Write documentations. (No user-facing SGLang documentation changes are required.)
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. (Accuracy results are above; no standalone SGLang speed benchmark was run.)
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ❌ Run #34307729743
Latest PR Test (Extra): ❌ Run #34307729362
Latest PR Test (AMD ROCm 7.2): ❌ Run #34307729485