Convert Arrow <--> Parquet, and hence Awkward <--> Parquet.

[jpivarski]

No description provided.

[jpivarski]

@martindurant As expected, Parquet support was rather easy—I just had to copy the old code. Awkward PartitionedArrays translate into row groups, top-level RecordArrays translate into record batches, and any other kind of array is presented as a record batch with one field whose name is the empty string. (This convention is also used when reading back: if there's only one field and its name is the empty string, we read back into a non-empty array.)

I also tested it on some ancient samples I made for OAMap, but even after all these years, most of the data structures are not supported by pyarrow:

https://github.com/scikit-hep/awkward-1.0/blob/97b165e53666d13a88da247ea18d577bfdf85761/tests/test_0341-parquet-reader-writer.py#L105-L246

To mitigate this, I added an explode_records option to ak.to_parquet. The transformation isn't lossy, but the result would have to be read back into ak.zip. Not ideal.

But hey, we have both eager and lazy reading now!