[DECOUPLED-MODE] ci: run the decoupled test gate in a real decoupled environment

[gulsumgudukbay]

Continues #4642, which I was asked to take over. I don't have push access to that branch (it lives in this repo, not a fork), so this comes from the ROCm fork and carries @shralex's commits plus one on top addressing the review comments.

Summary

• The gate now builds its own venv from the new generated_requirements/decoupled-requirements.txt and fails the step if
  google.cloud.storage, google.api_core, google.auth, gcsfs, jetstream or the diagnostics clients are importable in it. DECOUPLE_GCLOUD=TRUE only changes what gcloud_stub hands back, so a gate running in a venv that has google-cloud-storage installed would have passed Checkpoint_conversion: import GCS via gcloud_stub for decoupled mode #4003, [DECOUPLED-MODE] grpo_trainer: route goodput/vertex imports through gcloud_stub for decoupled #4180 and [DECOUPLED-MODE] checkpoint_conversion: import GCS via gcloud_stub in load_dynamic #4637.
• decoupled-requirements.txt is derived from the GPU pre-training lock minus the Google Cloud clients and the accelerator wheels, so the decoupled run uses the same versions as the gpu-unit job (checked against a gpu-unit run on main, no version differences) and installs on any hardware. A decoupled run on an accelerator adds the dropped entries back via --print-accelerator-requirements. Deriving it from a lock rather than from base requirements keeps regeneration offline, which is what lets the new decoupled-requirements pre-commit hook fail any requirements change that leaves it stale. Replaces requirements_decoupled_jax_0_7.1.txt, which nothing referenced.
• Step wiring fixes: it sat between the previous step's run: and its env:, so Run Tests lost all ten INPUTS_* variables; it ran in all four cpu-unit shards; and $PYTHON_EXE was empty across the step boundary.
• Drops decoupled_target from the integration test and from the files pytest.ini ignores, where the marker never took effect, and corrects the marker docs.

Prerequisite

#4637 has to land first. Without it, tests/unit/checkpointing_test.py cannot be collected once the Google Cloud packages are genuinely absent, and the whole runxits 2.

Test plan

• Ran the step's run: block as is under bash -e against a built wheel, with [DECOUPLED-MODE] checkpoint_conversion: import GCS via gcloud_stub in load_dynamic #4637 applied: 46 passed, 2 skipped, no collection errors.
• Guard negative cases: with google-cloud-storage present it exits 1 and stops the step before pytest; same for gcsfs; a missing wheel fails loudly rather than skipping.
• The same selection with DECOUPLE_GCLOUD unset fails at collection (ModuleNotFoundError: ml_goodput_measurement), so the gate can't pass trivially.
• pre-commit run passes on every touched file, including the new hook.

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.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!