[DECOUPLED-MODE] grpo_trainer: route goodput/vertex imports through gcloud_stub for decoupled

[gulsumgudukbay]

Description

The top-level from ml_goodput_measurement.src.goodput import GoodputRecorder and from maxtext.common.vertex_tensorboard import VertexTensorboardManager (which pulls in cloud_accelerator_diagnostics) hard-require cloud-only packages that are intentionally absent in decoupled environments (DECOUPLE_GCLOUD=TRUE). A single ModuleNotFoundError at import time aborted pytest collection for the whole suite (exit 2).

Resolve both through the existing decoupled-aware paths, mirroring pre_train/train.py:

• GoodputRecorder (used only as a type annotation) now comes from the stub-aware maxtext.common.goodput.goodput namespace.
• VertexTensorboardManager now comes from gcloud_stub.vertex_tensorboard_modules().
  No behavior change when the real libraries are installed; decoupled runs fall back to the existing no-op stubs. Fixes rocm-decoupled collection (786/1157 collected, 371 deselected, exit 0).

Tests

Downstream CI decoupled ROCm run.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!