feat(nemo-gym): add SMG router backend

[NolenLiang]

What does this PR do ?

Adds SMG 1.9.0 as a selectable NeMo Gym inference-router backend.

• generalizes the vLLM Router process from feat(nemo-gym): route rollouts through vLLM Router #3518 into a backend-selectable inference router
• maps each backend's launch module, supported policy names, and session-affinity header
• validates unsupported backend/policy combinations before startup
• routes NeMo Gym policy traffic through either vLLM Router or SMG without changing worker endpoints
• pins SMG 1.9.0 in the nemo_gym optional dependency set

This is a stacked PR whose base is #3518, so its diff contains only the SMG integration increment.

Issues

Depends on:

• feat(nemo-gym): route rollouts through vLLM Router #3518
• NVIDIA-NeMo/Gym#2347, which forwards session-affinity headers and consumes inline token IDs without requiring a router /tokenize endpoint

Usage

Configure the router block under nemo_gym:

router:
  enabled: true
  backend: smg
  policy: consistent_hash

Set backend: vllm_router to use the vLLM Router implementation through the same configuration shape. The shared consistent_hash value maps to SMG's consistent_hashing CLI policy when SMG is selected.

Before your PR is "Ready for review"

Pre checks:

• Read and followed the Contributor guidelines
• Added unit coverage for both router backends, policy validation, lifecycle handling, NeMo Gym wiring, and port allocation
• Ran the relevant unit and functional checks locally
• Kept this first integration PR minimal; repository documentation is intentionally not included

Additional Information

Validation completed:

• 16 relevant NeMo RL unit tests passed
• 4 relevant NeMo Gym tests passed
• Ruff check and format check passed
• Pyrefly passed with 0 errors across 219 modules
• uv lock --check passed with the repository's container-pinned UV 0.11.28
• an SMG 1.9.0 smoke test with two mock vLLM workers verified readiness, consistent session routing, distribution across both workers, inline token-ID pass-through, and graceful shutdown with exit code 0

[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.

[github-actions]

❌ Submodule Fast-Forward Check Failed

Check based on commit: d45a886 (PR #3536 from nliang/smg-router-nemo-gym)

❌ Submodules that need attention:

Gym: ❌ Commits have DIVERGED from a common ancestor
TARGET (nliang/vllm-router-nemo-gym branch): https://github.com/NVIDIA-NeMo/Gym/commits/a4e16ce67f9b72cb2d48e4728a40374d12949ef5/
CURRENT (PR #3536 from nliang/smg-router-nemo-gym): https://github.com/NVIDIA-NeMo/Gym/commits/550912f2e721e80c3d608c8b2963d3e243bd0837/

Please ensure all submodule commits are fast-forwards of the nliang/vllm-router-nemo-gym branch before merging.

[NolenLiang]

The earlier submodule failure on d45a886b is resolved.

• feat(nemo-gym): route rollouts through vLLM Router #3518 now points the Gym submodule to 550912f2.
• This PR was rebased onto the updated feat(nemo-gym): route rollouts through vLLM Router #3518 head (1ee6e4f11).
• The current head (318cdf575) no longer changes the Gym submodule.
• The current submodule fast-forward check passes.

[NolenLiang]

/ok to test 16ebc41

[NolenLiang]

/ok to test 7b8b7e0