feat: support R3 async RL without TQ

[zyzhou5]

What does this PR do ?

• Preserve R3/router-replay routed_experts when building policy input batches in GRPO training paths.
• Enable the async no-TQ path to consume rollout-recorded routes for prev-logprob and training.
• Add a clear guard for the unsupported async+TQ combination.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[zyzhou5]

/ok to test c4305cd

[zyzhou5]

/ok to test 9f2afe8

[zyzhou5]

/ok to test 9f2afe8

[zyzhou5]

/ok to test 40fa5d3

[ZhiyuLi-Nvidia]

Awesome change. LGTM!

[zyzhou5]

/ok to test 7a1c5ce

[zyzhou5]

/ok to test 3b617bf

[qiaochuz-nv]

Test coverage added:

1. test_grpo_router_replay_async_dp_off_daily_pr
   PR covered: NVIDIA-NeMo/RL #2908

   Functions / paths covered:
   - async GRPO R3 path without TransferQueue
   - data_plane.enabled=false
   - _build_async_grpo_train_data
     - verifies routed_experts survives async producer/consumer batching
     - verifies train/logprob data carries R3 route tensors
   - Megatron Qwen3-30B-A3B MoE policy path
   - TP4 / PP2 / CP2 / EP4 topology
   - non-colocated vLLM generation
   - RouterReplay replay in:
     - prev-logprob
     - train
   - CP2 token identity / zigzag route-sharding correctness
   - forward verifier matches_expected=true

   Status: PASS
   Evidence:
   - run_id: 20260706T185536Z-0496ec2b
   - SLURM job: 5579660
   - SLURM: COMPLETED|0:0
   - elapsed: 00:08:52
   - nodes: eos[0516,0518,0520]
   - train/loss steps: 2
   - R3 trace records: 21632
   - replay assignments:
     - prev-logprob: 3072
     - train: 3072
   - replay forward:
     - prev-logprob: 3072
     - train: 3072
   - replay backward train: 3072
   - forward verifier:
     - prev-logprob: 3072
     - train: 3072
   - CP token identity rows: 47168