Make Vertex AI evaluation optional to avoid litellm bloat

[Vamsi-klu]

Problem

apache-airflow-providers-google base depends on google-cloud-aiplatform[evaluation] which pulls litellm, scikit-learn, ruamel-yaml, tqdm, huggingface-hub, tokenizers. Users of GCS/BigQuery get ML eval stack they never use. litellm has frequent proxy-server CVEs (auth bypass, RCE) triaged by every user.

Root cause: providers/google/pyproject.toml:86
"google-cloud-aiplatform[evaluation]>=1.155.0"

Code needing evaluation: vertexai.preview.evaluation.EvalResult, EvalTask used only in GenerativeModelHook.get_eval_task() and run_evaluation() and RunEvaluationOperator.

What I did

Part A: pyproject.toml

• Change base from "google-cloud-aiplatform[evaluation]>=1.155.0" to "google-cloud-aiplatform>=1.155.0"
• Add optional extra "evaluation" = ["google-cloud-aiplatform[evaluation]>=1.155.0"]

Part B: Lazy import guard in hook

• Try/except ImportError for EvalResult/EvalTask, store _evaluation_import_error
• In get_eval_task() and run_evaluation(), raise AirflowOptionalProviderFeatureException with install instruction if import failed
• Operator file doesn't import EvalTask at top-level, hook guard enough

Part C: Tests use pytest.importorskip("vertexai.preview.evaluation") to skip gracefully when extra not installed

Part D: Changelog breaking change note directly edited per provider rules

Generated files via prek update-providers-dependencies --all-files: README, docs/index.rst, provider info

Impact

• Base install no longer pulls litellm/scikit-learn (security + image size + dep hell improvement)
• Users needing RunEvaluationOperator must install apache-airflow-providers-google[evaluation]
• Breaking change for Eval users, mitigated via changelog and clear error message
• pandas remains base dep, ray remains

Testing

• python -c "from airflow.providers.google.cloud.hooks.gcs import GCSHook" works without evaluation extra
• Hook.get_eval_task() raises AirflowOptionalProviderFeatureException with helpful msg when mocked missing
• uv run --project providers-google --extra evaluation pytest ... -k evaluation → 1 passed
• ruff format/check passed
• Note: breeze ci-image rebuild needed

Fixes: #69323

Was generative AI tooling used to co-author this PR?

• Yes — Muse Spark 1.1
  Generated-by: Muse Spark 1.1

[Vamsi-klu]

Reviewers: @shahar1 (google provider CODEOWNER). Breaking change: evaluation extra now optional to avoid litellm bloat. Changelog entry added, provider docs regenerated via update-providers-dependencies.

Fixes #69323

---

Drafted-by: Muse Spark 1.1; reviewed by @Vamsi-klu before posting