Correctly update metadata to preserve features when concatenating datasets with axis=1

[mariosasko]

This PR correctly updates metadata to preserve higher-level feature types (e.g. ClassLabel) in datasets.concatenate_datasets when axis=1. Previously, we would delete the feature metadata in datasets.concatenate_datasets if axis=1 and restore the feature types from the arrow table schema in Dataset.__init__. However, this approach only works for simple feature types (e.g. Value).

Fixes #3111

[albertvillanova]

Thanks @mariosasko. And also thanks for the test!