Make read_parquet_dataset() less I/O intense

[hombit]

Currently hipscat.io.file_io.read_parquet_dataset() makes some assumptions which could be suboptimal for some use-cases. For instance, it sets pyarrow.dataset.dataset(exclude_invalid_files=True) which force-scan the whole catalog. I propose to make this behavior optional.

This change would also require updating ignore_prefixes, which currently doesn't include JSON and FITS metadata files, which makes pyarrow.dataset.datase believe that they are parquet files. Probably, we could also allow user to provide custom ignore_prefixes.

The probable solution would be adding **kwargs to the function and allow to pass any arguments to pyarrow.dataset.dataset arguments.