fix: set TIS type for Automodel GRPO release recipes

[zpqiu]

What does this PR do ?

Fixes Automodel GRPO release recipes that can crash when truncated_importance_sampling_ratio is configured without an explicit truncated_importance_sampling_type.

The Qwen3.5 DAPO and GLM-4.7 Flash Automodel recipes already set truncated_importance_sampling_ratio: 2, but did not set the corresponding sampling type. This PR sets loss_fn.truncated_importance_sampling_type: tis in both recipes so the ratio is interpreted through the intended TIS path.

Additional release-test stabilization:

• Keeps the GLM release test median train/token_mult_prob_error check.
• Removes only the GLM step-30 point check for train/token_mult_prob_error, which has high tail variance.
• Keeps train/gen_kl_error as the policy/generation mismatch check.

Issues

N/A

Usage

No user-facing usage change. This updates release recipe/test configuration only.

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

Validation performed:

• bash -n tests/test_suites/llm/grpo-glm47-flash-4n8g-automodel.sh
• Qwen3.5 DAPO Automodel release rerun passed after setting truncated_importance_sampling_type: tis.
• GLM-4.7 Flash Automodel rerun completed 30 steps. The retained checks passed:
  • median(train/token_mult_prob_error) = 1.059859812259674 < 1.1
  • mean(train/gen_kl_error) = 0.007180878799408675 < 0.01
  • train/reward[30] = 0.666015625 > 0.3
  • max(validation/accuracy) = 0.5291666388511658 > 0.2
  • mean(train/grad_norm) = 0.15234075884024303 < 0.6
• The removed GLM single-step check was train/token_mult_prob_error[30] < 1.1; the rerun observed 17.038604736328125 at step 30 while the median and KL checks remained healthy.

[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.