[feat] Add Qwen3 MoE true-on-policy SP/PP support

[maocheng23]

Based on maocheng23#4 and retargeted to sgl-project/sglang:sglang-miles.

Summary

Adds the SGLang rollout-side contract and deterministic kernel plumbing for Qwen3-30B-A3B MoE true-on-policy training, including the SP+PP contract update used by the coupled Miles and Megatron-LM PRs.

This PR is one of three coupled PRs that should be reviewed and landed together because they extend the qwen3_moe_true_on_policy_v1 contract to allow Megatron sequence parallel and pipeline parallel training while preserving exact SGLang rollout parity.

Companion PRs:

• Miles: [feat] Add SP and PP support for qwen_moe true on policy radixark/miles#1088
• Megatron-LM: [feat] Add SP and PP kernels for qwen_moe true on policy radixark/Megatron-LM#31

Main Changes

• Add Qwen3 MoE true-on-policy runtime policy/schema wiring under python/sglang/srt/true_on_policy/.
• Add batch-invariant and TP-invariant ops used by deterministic rollout/recompute parity paths.
• Route Qwen3 MoE, layernorm, communicator, model runner, CUDA graph, and TP worker paths through the true-on-policy runtime policy where needed.
• Disable NVLS on the deterministic TP decode path so TP reductions match the Megatron parity contract.
• Remove earlier debug/weight-audit and legacy true-on-policy target paths from the final branch.
• Flip the shared schema flag so Megatron sequence parallel is allowed for the dense and MoE true-on-policy schemas.

Validation

Cleaned 8-GPU ion7 real-workload run:

• Run id: moe_sppp_pr_tp2_pp2_ep2_onpolicy_real3_cleaned_noci_nosave_260524_ion7
• Ray job: raysubmit_a9NCtiu6K5dRVCsc
• Topology: Megatron TP=2, PP=2, EP=2, ETP=1, sequence parallel enabled; SGLang 4 rollout engines, each TP=2, EP=2.
• Result: Ray job succeeded, all 8 GPUs returned to 0 MiB.
• Step 1: train/train_rollout_logprob_abs_diff=0.0, train/train_rollout_kl=0.0, grad_norm=0.0377418305, weight version 1.0, mixed version 0.0.
• Step 2: train/train_rollout_logprob_abs_diff=0.0, train/train_rollout_kl=0.0, grad_norm=0.0389085777, weight version 2.0, mixed version 0.0.
• Timing: cleaned step times 485.2571s and 462.8948s, about 2.6-2.9% slower than the previous on-policy no-debug timing run and in the same overhead band versus off-policy.

Focused checks after cleanup:

• git diff --check
• Miles fast checks: 53 passed
• Megatron targeted extension/MoE checks: 3 passed
• Megatron 8-rank tensor-parallel mapping check passed on all ranks
• Full SP+PP 8-GPU E2E exact-logprob rerun on ion7

Local record:
recovery/qwen3_moe_sppp_clean/journal/2026-05-23-sppp-e2e.md

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.