[GLM-5.3] Remove experimental DSA metadata optimizations

[Fridge003]

Motivation

Stacked on #36507 (xinyuan/glm-5.3-flash-support, base d769a1d265b8c732b72b946b6ba2a16ae1d85815). Remove the experimental metadata optimizations while retaining GLM-5.3 Flash's No-RoPE, KPool, EAGLE, and TileLang/FA3 fallback support.

This should be added back after clean up.

Modifications

• Remove the runtime flags and all their feature paths:
  • SGLANG_EXPERIMENTAL_DSA_KPOOL_METADATA_FUSION
  • SGLANG_EXPERIMENTAL_DSA_INGRAPH_VERIFY_METADATA
  • SGLANG_EXPERIMENTAL_DSA_INGRAPH_VERIFY_METADATA_DG_OUT_OF_GRAPH
• Remove fused KPool metadata generation, draft-step sibling metadata reuse, in-graph verify state/capture/replay, and the out-of-graph DeepGEMM residual helper. KPool decode/verify/draft-extend always use the ordinary pool-aware metadata path.
• Remove the experimental fused-kernel bounded-scan/launch-preparation changes and preallocated top-k-plan API. Retain the existing non-KPool fused kernels and 64-bit page-table output addressing (the latter prevents large-context offset overflow).
• Restore ordinary out-of-graph top-k-plan refresh. Preserve per-backend DeepGEMM schedule/KPool refresh, multi-step precompute, EAGLE's verify hook, and static-buffer address stability.
• Delete the three environment variables outright: no aliases, deprecated-environment entries, or warning-only tests remain.
• Add a GPU regression test for decode, target-verify, draft-extend-v2 and MTP-precomputed metadata, with actual KPool kernels, DeepGEMM/top-k plans and CUDA graph replay. Checks pool/page boundaries, live tails beyond top-k, request remapping, accept lengths, refreshed schedule/plan contents and stable buffer addresses.

dsa_backend.py is reduced by 480 lines. No changes to the model forward math or No-RoPE/TRTLLM/TileLang/FA3 dispatch paths.

Accuracy Tests

GPU validation at 62f4493387 (before the follow-up deletion of the warning-only registry entries/test), on temporary rx devbox codex-glm53-metadata-trim: 1x NVIDIA H200, driver 580.105.08, PyTorch 2.13.0+cu130.

PYTHONPATH=python python -m pytest -q \
  test/registered/unit/test_environ.py \
  test/registered/unit/layers/attention/test_dsa_kpool_metadata.py \
  test/registered/kernels/test_dsa_kpool_multi_pool.py \
  test/registered/kernels/ops/attention/test_dsa_metadata.py
# 27 passed, 14 subtests passed

PYTHONPATH=python python -m pytest -q \
  test/registered/kernels/ops/attention/test_topk_v2.py -k '8-4096'
# 9 passed, 269 deselected

Additional checks:

• Ran the new metadata test against unmodified base GLM-5.3-Flash support #36507 on the same H200, with the bare-backend fixture's three former options set to false: both tests/all modes pass. The ordinary-path behavior is retained.
• Follow-up removal: repository-wide search finds no references to any of the three flags. A stdlib-only import smoke test confirms that setting them neither registers an option nor emits deprecation warnings. The existing environment unit-test file is unchanged relative to GLM-5.3-Flash support #36507.
• AST comparison confirms No-RoPE/TRTLLM/FA3/TileLang forward methods, the core EAGLE verify hook, KPool init/capture/fallback methods are unchanged. Ordinary precomputed replay is AST-equivalent after inlining the former shared-copy helper, including schedule/top-k/KPool refreshes.
• Changed-file pre-commit checks and CI registration validation pass.

Full GLM model/GSM8K e2e was not rerun for this removal. No full CI or run-ci was requested; validation was confined to the devbox tests above.

Speed Tests and Profiling

No new speed benchmark. This intentionally removes optional acceleration: workloads that enabled these options lose those optimizations, and reverting the shared bounded-scan/preallocated-plan changes can increase metadata overhead (especially large-context non-KPool workloads). Accuracy thresholds and serving configurations are unchanged.

Checklist

• Format code with pre-commit.
• Add focused unit/kernel regression coverage and CI registration.
• Document the outright removal in this PR; no deprecation compatibility layer.
• Follow SGLang code style.
• Full-model accuracy/speed benchmark (not run; focused devbox validation only).

---

CI States

Latest PR Test (Base): ❌ Run #33928287001
Latest PR Test (Extra): ❌ Run #33928286867
Latest PR Test (AMD ROCm 7.2): ❌ Run #33928287067