Improve Uproot transformer serialization

[kyungeonchoi]

Story

Current uproot transformer performs the following steps for the serialization of data:

1. Awkward Table returned from generated_transformer.py
2. Convert Awkward Table to Pandas DataFrame (DF)
3. Convert Pandas DF to Arrow
4. Save Arrow as Parquet or stream to Kafka

Remove unnecessary step(s) and implement the new method suggested by Jim Pivarski to improve the performance.

Acceptance Criteria

1. Better performance of Uproot transformer

[kyungeonchoi]

Awkward supports a direct conversion from awkward.Table to Arrow (awkward.toarrow), but the layout of awkward.Table returned by generated_transformer.py is not compatible with this function. It has to be top-level ChunkedArrays to be converted to Arrow.

Two solutions:

1. Immediate: convert awkward.Table to awkward1.Table and then back to awkward.Table. This fixes the layout issue of the original awkward.Table. Now awkward.toarrow should work. awkward1 >= 0.2.19 supports this method.
2. Near future: Uproot will produce awkward1.Table and awkward1 will produce Arrow (toarrow and fromarrow #68 scikit-hep/awkward#224)

Quick benchmark of Awkward Table to Arrow serialization for the 9904 (row) x 20 (column) Awkward table (tested at UC River):

• Current: 7.95 sec
• Immediate solution: 0.01 sec

Significant improvement is expected for large tables.

[jpivarski]

Although it's technically closed, work on scikit-hep/awkward#224 hasn't stopped. I've taken over for Anish and I'm touching up his work as scikit-hep/awkward#263. It should be ready pretty soon—these are the final dottings and crossings of i's and t's.