Google provider base-depends on google-cloud-aiplatform[evaluation], force-installing litellm/scikit-learn for every user

**Repo:** apache/airflow · **Labels:** kind:bug, provider:google

---

### Title
Google provider base-depends on `google-cloud-aiplatform[evaluation]`, force-installing litellm / scikit-learn for every user

### Apache Airflow Provider(s) version
`apache-airflow-providers-google` 22.0.0 — also confirmed on latest **22.2.0**.

### What happened
`apache-airflow-providers-google` declares `google-cloud-aiplatform[evaluation]` as an **unconditional base dependency** (no extra marker):

- 22.0.0: `google-cloud-aiplatform[evaluation]>=1.145.0`
- 22.2.0: `google-cloud-aiplatform[evaluation]>=1.155.0`

The `[evaluation]` extra transitively installs a full ML/eval stack that has nothing to do with the core Google hooks/operators:

```
apache-airflow-providers-google
└── google-cloud-aiplatform[evaluation]
├── litellm → huggingface-hub → tokenizers → tqdm
├── scikit-learn
└── ruamel-yaml
```

Anyone who installs the provider only for, say, `GCSHook` / `BigQueryHook` still gets `litellm`, `scikit-learn`, and friends.

### What you think should happen instead
The Vertex AI **evaluation** feature set should be opt-in, not forced on all provider users. Either:
1. Move it behind a provider extra (e.g. `apache-airflow-providers-google[vertex-eval]`), or
2. Base-depend on `google-cloud-aiplatform` **without** the `[evaluation]` extra (users who need eval add it themselves).

### Why it matters
- **Image-size bloat**: litellm + scikit-learn + the HF/tokenizers chain is a large install for users who never touch Vertex evaluation.
- **CVE noise**: `litellm` ships a steady stream of *proxy-server* CVEs (e.g. auth bypass, sandbox RCE, privilege escalation) that surface in `pip-audit` for every provider user, even when litellm is never imported and its proxy is never run — each one has to be triaged/ignored downstream.

### How to reproduce
```bash
pip install apache-airflow-providers-google==22.2.0
pip show litellm scikit-learn # both present, transitively via the [evaluation] extra
```

### Anything else
The `evaluation` extra of `google-cloud-aiplatform` is the sole source of `litellm`/`scikit-learn`/`ruamel-yaml` in a provider install. Making it optional would let downstreams that use `google-genai` (the newer unified SDK) for Vertex AI avoid the old `aiplatform` eval SDK entirely.

Start with the Google provider’s dependency declaration and any packaging or dependency tests; the issue does not name their paths. Reproduce the current install with the provided pip commands, then trace how the evaluation dependency is exposed. Done means a base provider install no longer pulls the evaluation stack, while the chosen evaluation behavior remains covered by the project’s dependency checks.

Written by the indexing model from the issue text.