Add CritPt Benchmark

[fsiino-nvidia]

Fixes DCO issue in #1537

--

Migrates the CritPt benchmark from nemo-skills into nemo-gym. This keeps the workflow gym-shaped (per-rollout reward flowing through the standard /verify contract).

How it works in gym:

• Architecture - Single phase: agent runs per-rollout /run -> /verify.

• Batching - Resources server buffers concurrent verify() calls until 70 unique problem_ids accumulate, then fires once and resolves all 70 awaiting futures with the same aggregate accuracy

• num_repeats > 1 - Each repeat of a problem_id opens a new pending batch. N repeats -> N independent AA API calls.

• Reward - Same as skills in that the aggregate is distributed to all batch members as their reward.

• Headline metric - pass@k/accuracy via compute_metrics override. Numerically equal to nemo-skills' accuracy at num_repeats=1.

Test results:
With ultra mopd/step36 checkpoint:

Key metrics for critpt_benchmark_agent:
{
    "mean/reward": 0.05714285714285713,
    "mean/accuracy": 0.05714285714285713,
    "mean/timeout_rate": 0.0,
    "mean/input_tokens": 6107.028571428571,
    "mean/output_tokens": 2514.1857142857143,
    "mean/total_tokens": 8621.214285714286
}

wandb: https://wandb.ai/nvidia/fsiino-gym-dev/runs/lle631hq

[cmunley1]

would be good to reproduce some other models eg kimi, minimax, but lgtm

[gwarmstrong]

small request for additional fields

[gwarmstrong]

can you add the license?

[gwarmstrong]

also description and value?

[fsiino-nvidia]

Added all 3 in 719cd07 . Thanks!