Allow postponing dataset integrity checks in `NextGenHDFDataset`

[NeoLegends]

Follow-up from #1315, where @JackTemaki noticed the NextGenHDFDataset goes through the entire data on startup to perform integrity checks -- and indeed RETURNN startup w/ that dataset is quite slow. This PR allows delaying these checks to training time, saving startup time at the cost of detecting data errors only later on.

[albertz]

Why do you put "chore:" into the description? Where do you get this from?

This is inconsistent to our normal commit messages, so we should not use this.

[albertz]

Does this really need to be a new option? I personally prefer to have fewer options if possible, esp if not really needed. Can't we just always do those integrity checks lazily?

[albertz]

Please check failing tests.

  returnn/datasets/hdf.py:548: WEAK WARNING PyIncorrectDocstringInspection: Missing parameter data in docstring
  returnn/datasets/hdf.py:548: WEAK WARNING PyIncorrectDocstringInspection: Missing parameter seq_name in docstring
  returnn/datasets/hdf.py:550: WARNING PyUnresolvedReferencesInspection: Function 'check_data_integrity' does not have a parameter 'numpy'
  returnn/datasets/hdf.py:550: WEAK WARNING PyIncorrectDocstringInspection: Unexpected parameter numpy in docstring
  returnn/datasets/hdf.py:551: WARNING PyUnresolvedReferencesInspection: Function 'check_data_integrity' does not have a parameter 'str'
  returnn/datasets/hdf.py:551: WEAK WARNING PyIncorrectDocstringInspection: Unexpected parameter str in docstring
  returnn/datasets/hdf.py:603: WEAK WARNING PyIncorrectDocstringInspection: Missing parameter data in docstring
  returnn/datasets/hdf.py:603: WEAK WARNING PyIncorrectDocstringInspection: Missing parameter seq_name in docstring
  returnn/datasets/hdf.py:605: WARNING PyUnresolvedReferencesInspection: Function 'check_data_integrity' does not have a parameter 'numpy'
  returnn/datasets/hdf.py:605: WEAK WARNING PyIncorrectDocstringInspection: Unexpected parameter numpy in docstring
  returnn/datasets/hdf.py:606: WARNING PyUnresolvedReferencesInspection: Function 'check_data_integrity' does not have a parameter 'str'
  returnn/datasets/hdf.py:606: WEAK WARNING PyIncorrectDocstringInspection: Unexpected parameter str in docstring
  returnn/datasets/hdf.py:668: WEAK WARNING PyIncorrectDocstringInspection: Missing parameter data in docstring
  returnn/datasets/hdf.py:668: WEAK WARNING PyIncorrectDocstringInspection: Missing parameter seq_name in docstring
  returnn/datasets/hdf.py:670: WARNING PyUnresolvedReferencesInspection: Function 'check_data_integrity' does not have a parameter 'numpy'
  returnn/datasets/hdf.py:670: WEAK WARNING PyIncorrectDocstringInspection: Unexpected parameter numpy in docstring
  returnn/datasets/hdf.py:671: WARNING PyUnresolvedReferencesInspection: Function 'check_data_integrity' does not have a parameter 'str'
  returnn/datasets/hdf.py:671: WEAK WARNING PyIncorrectDocstringInspection: Unexpected parameter str in docstring

Please use PyCharm to directly see those inspections. In general, be consistent to other code.

[albertz]

I guess @JackTemaki and/or @patrick-wilken should otherwise review this.

[NeoLegends]

Can't we just always do those integrity checks lazily?

I don't really have an opinion for or against that here -- it's mainly detecting data format errors early vs. saving time. If you're sure your data is good there is no reason to do these checks eagerly.