Fix nested infos decoding in ArrowStorage

[yurekami]

_decode_info in arrow_storage.py checks isinstance(field, pa.StructArray), but field is a pa.Field (it comes from iterating the struct type), so the branch that recurses into nested dictionaries never runs. Nested info dictionaries stored with the arrow/parquet formats come back as an object array of per-step dicts, with the inner arrays flattened, instead of the nested dict of stacked arrays that the HDF5 backend returns:

hdf5     infos['nested'] -> dict: {'matrix': ('float64', (4, 2, 2))}
arrow    infos['nested'] -> ndarray: object (4,) [{'matrix': array([0., 0., 0., 0.])} ...]

This PR checks the field type with pa.types.is_struct so nested dictionaries are decoded recursively (matching _encode_info), and adds test_add_episodes_with_nested_infos in tests/dataset/test_minari_storage.py, parametrized over all storage formats. The test fails on arrow and parquet before the change and passes after.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have run the pre-commit checks with pre-commit run --all-files
• I have run pytest -v and no errors are present.
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

🤖 Generated with Claude Code

https://claude.ai/code/session_011vZKTbRaDh8FuXSbVpQomD