Refactor HDF5 and preserve tree structure

[klamike]

Closes #7741. Followup to #7690

• Recursive parsing and feature inference, to preserve the tree structure of the file. Note this means we now visit all links in the file. It also means we have to call combine_chunks on any large non-root datasets.
• Support for complex64 (two float32s, used to be converted to two float64s)
• Support for ndim complex, compound, more field types for compound (due to reusing the main parser, compound types are treated like groups)
• Cleaned up varlen support
• Always do feature inference and always cast to features (used to cast to schema)
• Updated tests to use load_dataset instead of internal APIs
• Removed columns in config. Have to give Features (i.e., must specify types) if filtering

[klamike]

Had to add this when trying a large dataset.. this is because all groups are represented by StructArray now, which needs non-chunked arrays when constructing

[klamike]

@lhoestq this is ready for you now!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[lhoestq]

LGTM ! 🙌