Use pyarrow to write parquet files

[jaychia]

Addresses: #123

On SCALE_FACTOR=10:

• Lineitem Parquet file before the change have 20k rowgroups
• Lineitem Parquet file after the change only has 57 rowgroups

This does affect benchmark results quite a bit, depending on how resilient the Parquet reader implementations are to these poorly written Parquet files. However I think for the sake of having a benchmark that represents the expected/common case we should write the Parquet files properly!

[ritchie46]

I think we should fix the culprit upstream instead of bandaid it here.

[jaychia]

That makes sense @ritchie46, I'll close this PR in favor of an upstream fix