Decouple GCloud dependencies

[gulsumgudukbay]

Description

Enable running the test suite completely offline (no gs://, no gcloud, no Vertex AI) when the environment variable: DECOUPLE_GCLOUD=TRUE

Key Environment Variables
DECOUPLE_GCLOUD=TRUE: Primary toggle for offline/"decouple from gcloud" mode.
LOCAL_BASE_OUTPUT: Optional override for local run/output directory (defaulted to decoupled_datasets/gcloud_decoupled_test_logs in most tests).
LOCAL_GCLOUD_PROJECT: Placeholder to satisfy code paths expecting a project (set in CI).

A minimal in-repo dataset directory (i.e. decoupled_datasets) is used instead of:
gs://maxtext-dataset (TFDS / ArrayRecord)
gs://maxtext-dataset/hf/... (Parquet)

In decoupled mode:

• All tests and helper scripts redirect base_output_directory to decoupled_datasets/gcloud_decoupled_test_logs and modify input dataset accesses to decoupled_datasets if DECOUPLE_GCLOUD=TRUE.
• vertex_tensorboard no-op in decoupled mode.
• max_utils.get_project() returns local placeholder without gcloud invocation when decoupled.
• Decoupling script added (decouple.py) with stubs for JetStream, Tunix, and diagnostics, so that if we are in decoupled mode, these stubs are used and GCE-dependent imports are cancelled earlier.
• Serving/JetStream/vLLM entrypoints fail-fast when decoupled so that training continues with safe fallbacks.
• Pytest markers created (external_serving, external_training, tpu_only). If a test has any of these markers, it is marked as decoupled, so that the end user can run the GCE-independent tests using -m decoupled marker.

Tests

In progress...

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.