Support loading strings with the pyarrow backend to use the string[pyarrow] pandas dtype

[smcguire-cmu]

Currently for string columns, pandas will load the strings as native python strings, and dask will then create a slow task to convert them all to pyarrow strings. Pandas has recently introduced support for the pyarrow string dtype, and can load strings from parquet files directly into a pandas df with the pyarrow string type by specifying dtype_backend="pyarrow" as an option in the pd.read_parquet call.

We support passing kwargs to this function, but when generating the dask meta DataFrame from the parquet schema, we don't use pyarrow string types, and so we get a meta mismatch. So this needs to be updated here, and tested that the new dtype works with the other from_delayed functions for operations like crossmatching and joining where we generate the meta.