feat: length-aware rollout scheduling for cross-DP generation

[zaifengp-maker]

What does this PR do ?

Schedules rollout requests by predicted output length instead of FCFS, so a cross-DP generation wave is not held open by one long straggler.

Within a wave, requests are admitted and placed across DP replicas by a length
estimate rather than arrival order. The PR also carries the tracing and
profiling used to evaluate the idea.

• nemo_rl/models/generation/vllm/lfs/ — the admission and cross-DP dispatch
  package: scheduler, dispatcher, per-engine schedulers, concurrency cap,
  group state, validation, diagnostics.
• vllm_step_trace.py — per-step engine trace.
• model_step_gpu_profiler.py — bounded Nsight capture around model steps.
• nested_runtime_env.py / ray_executor.py — propagate a runtime env into
  vLLM's Ray workers without mutating vLLM sources.
• inflight_profiler.py, vllm_metric_sampler.py — in-flight rollout and
  DP-batch statistics.
• forced_sequence_logits_processor.py — deterministic sequence replay, used
  to make scheduling experiments reproducible.

plus the corresponding changes in vllm_generation, vllm_worker,
vllm_worker_async, vllm_backend, grpo, rollouts, async_utils and
dapo_math, and 19 unit tests.

Issues

None filed yet. Happy to open one to hold the design discussion if that is
preferred over discussing it here.

Usage

LFS is off by default; generation behavior is unchanged unless it is enabled.
It is currently driven by environment variables rather than config keys:

# Cross-DP admission mode. Off unless set.
#   fcfs | lfs | predicted_lfs | history_lfs | oracle_probe_lfs | exact_length_lpt
NRL_VLLM_CROSS_DP_SCHED=lfs

# Single-engine length-grouped admission (independent of the cross-DP path)
NRL_VLLM_LFS_SCHED=1
NRL_VLLM_LFS_SCHED_G=16          # group size

The one config key added is for the tracing facility, not the scheduler:

policy:
  generation:
    vllm_cfg:
      enable_vllm_step_trace: true   # read-only per-step batch composition

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests? — 19 unit tests under tests/unit/models/generation/vllm/
• Did you run the unit tests and functional tests locally? — no. Not run against this branch.
• Did you add or update any necessary documentation? — no. Not written yet, pending direction on whether this belongs in tree.

[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.