ci: update gpt3_7b_tp4_pp1_memory_speed gb200 golden values

[ko3n1g]

After https://github.com/NVIDIA/Megatron-LM/commits/main/tests/functional_tests/shell_test_utils/_run_training.sh we forgot to update the nightly values of the test in question. Perhaps the test was still in queue at the time we updated values. Looking at the dashboard, its the only change that happen in the span that clear explains a shift in golden values.

Claude summary

Summary

Heals the deterministic golden-values regression on gpt3_7b_tp4_pp1_memory_speed (dgx_gb200, dev environment) introduced by #3779 (rampup batch size scheduler replacement).

• The APPROXIMATE test already passes — only the EXACT (deterministic) comparison failed.
• Deltas are last-decimal noise on lm loss and num-zeros; no metric drift on iteration-time, mem-allocated-bytes, or mem-max-allocated-bytes.
• Other env/platform combos (a100, h100, lts) are unaffected.

Verification

• Failing job (pre-fix): https://gitlab-master.nvidia.com/ADLR/megatron-lm/-/jobs/302065814
• Healing pipeline (all green, includes nightly run on gb200): https://gitlab-master.nvidia.com/ADLR/megatron-lm/-/pipelines/50196728
• Healed test job: https://gitlab-master.nvidia.com/ADLR/megatron-lm/-/jobs/310457478

Example diff

-            "3": 12.60284,
+            "3": 12.60283,
-            "25": 10.99869
+            "25": 10.99872

[ko3n1g]

/ok to test d23d965

[github-actions]

This PR has been automatically converted to draft because all PRs must start as drafts.

When you are ready for review, click Ready for Review to begin the review process. This will:

1. Add the oncall reviewer (optional reviewer)
2. Add required review teams based on your changes

See the contribution guide for more details.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.