feat: verifiers / environments hub integration

[cmunley1]

enables using environments hub envs in NeMo Gym with NeMo RL for training.

#446

[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.

[cmunley1]

ascii-tree

[cmunley1]

acereason-math

[cmunley1]

i3-math

[cmunley1]

multi turn seems to work, if we disable monotonicity checks and ensure consistent logprob dtype in nemo rl

[ahmadki]

tests would also be appreciated

[cmunley1]

would like for someone from prime intellect to take a look, and also test more environments to provide a longer list of what is working today (seems to be various version mismatches or other issues with some envs).

however, I think this is good to merge. Can always open another PR for more verified environments, or based on PI feedback

@bxyu-nvidia @cwing-nvidia

[cmunley1]

note that multi-turn does not support on-policy token-id correction with these envs, so training requires disabling this assert. In tests on alphabet-sort with it disabled training seems okay but we need to ensure the token ids look fine without this assert, or add support to use replace_prefix_tokens in nemo rl

https://docs.nvidia.com/nemo/gym/latest/contribute/rl-framework-integration/openai-compatible-http-server-on-policy-correction.html#on-policy-token-id-fix

maybe I am missing something in verifiers, e.g. PrimeIntellect-ai/verifiers#626