feat(model_loader): revive weight loader v2 dense tranche 1 (PR2 of #31051)

[joeqth]

Motivation

The weight loader v2 migration plan (#31051) stalled after PR1 (#28671) was
merged: the PR2–PR6 drafts (#31264–#31268) have been closed with no activity
since July. This PR revives the first PR2 tranche so the migration can
continue.

Full credit for the implementation goes to @JD-ETH — this PR rebases their
draft #31264 (commit f64a164) onto current main.

Modifications

• model_loader/auto_loader.py: add load_with_stacked_dispatch helper used
  by module-level MLP/Attention loaders
• Opt-in v2 dense loaders (behind SGLANG_ENABLE_WEIGHT_LOADER_V2, default
  off) for qwen3, gemma2, glm4, granite, internlm2, olmo, olmo2
  and the qwen3_classification mandatory wrapper — same env-gated dual-path
  shape as llama/qwen2 from PR1
• Legacy paths preserved verbatim as _legacy_load_weights
• Tests: CPU stacked-dispatch unit test, Qwen3-0.6B v1/v2 state-dict
  equivalence, e2e generation-match extension

Rebase note: the intervening refactor wave (#41816 make_pp_layers, #41750
boundary renames) did not touch any load_weights region; the three-way
merge resolved cleanly. Every order-sensitive legacy branch (tie-embedding
copies, prefix fixes, PP layer filtering, kv-scale remap) is preserved in
the v2 paths — e.g. qwen3's tie copy keeps the is_last_rank guard and the
stacked-parameter ordering is unchanged.

Accuracy Tests

• With the flag off (default) the legacy paths are byte-identical renames,
  so model outputs are unaffected.
• Env-on equivalence: test_qwen3_v1_v2_state_dict_identical (Qwen3-0.6B)
  compares the full state dict between V1 and V2 with rtol=0, atol=0.
• E2E: test_weight_loader_v2_e2e.py extension compares generation outputs
  V1 vs V2.
• CPU mapping unit test: test/registered/unit/model_loader/test_stacked_params_dispatch.py.
• I could not run GPU tests locally; CI will exercise the equiv/e2e tests.

Speed Tests and Profiling

Not applicable — weight loading only, no forward-path changes.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to the Write documentations. (none needed — the env flag is documented in Weight Loader v2 — protocol and migration plan (PR1–PR6) #31051)
• Provide accuracy and speed benchmark results according to the Test the accuracy and Benchmark the speed. (equivalence tests included; benchmarks N/A)
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #36873446858
Latest PR Test (Extra): ❌ Run #36873446450
Latest PR Test (AMD ROCm 10): ❌ Run #36873446674