feat: add NCCL timeout config, stale ZMQ socket cleanup, OmegaConfig resolvers

[puneeshkhanna]

What does this PR do ?

• Add configurable NCCL collective timeout via NRL_DIST_TIMEOUT_MINUTES env var (default 10 mins) to handle slow cross-node collectives during HF->Megatron reshard and policy->vLLM refit.
• Self-heal stale IPC socket files in maybe_init_zmq() to prevent EADDRINUSE errors after unclean exits (crash, kill -9, OOM).
• Register "add" and "sub" OmegaConf resolvers for use in config files. Helps to have below auto calculations in config files for agent harnesses.
  max_input_tokens: ${sub:${policy.generation.vllm_cfg.max_model_len}, ${policy.generation.max_new_tokens}}
  max_output_tokens: ${policy.generation.max_new_tokens}

Issues

List issues that this PR closes (syntax):

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

New tests not required.

/review-pr comments below:
Files changed: 3

Findings (scored ≥80)

None. LGTM.

Filtered (scored <80)

All candidates were either consistent with codebase conventions (env var ValueError — 92% confidence it's a non-issue) or pre-existing gaps (missing resolver tests). No
new bugs or guideline violations introduced by this diff.

The changes are clean and ready.

[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.