Skip notebook pip installs when Kaggle files are cached

A Databricks serverless notebook that starts with '%pip install kagglehub' fails the whole job whenever PyPI is unreachable, even though the data is already cached in a UC volume. kagglehub 1.x caches at $KAGGLEHUB_CACHE/datasets/<owner>/<name>/versions/<n>/<file>. Check those paths first and return them without importing kagglehub. Only on a cache miss, run subprocess pip install, then importlib.invalidate_caches(), then import. Unit-test the cached branch with monkeypatch.setitem(sys.modules, 'kagglehub', None), which makes any import raise.