fix: token mult prob error plot masking

[1ytic]

What does this PR do ?

Fixes the token_mult_prob_error debug plot so it does not select fully masked samples and labels the recomputed policy logprobs accurately.

Issues

N/A

Usage

N/A. This is a logging/debug-plot fix.

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

The plot previously computed the per-sequence multiplicative probability error after applying sample_mask, but still allowed rows with zero valid tokens to participate in argmax. Fully masked rows therefore produced 0 / 0 = nan, and the plot could show token_mult_prob_error=nan.

The orange logprob line was also labeled as reference policy, but the data comes from prev_logprobs, i.e. the training policy recomputation used for GRPO's behavior-policy comparison, not the frozen reference policy.

Before fix, the debug plot could select a fully masked sequence and show both symptoms:

After fix, the plot selects an unmasked sequence, reports a finite token_mult_prob_error, and labels the orange line as the training policy recompute:

[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.