Add intelligent sampling in ParquetSource

[dsmilkov]

The parquet reader can now read from local files, S3 or GCS. If the dataset is sharded, the reader takes a glob pattern to load multiple files.

The reader now takes shuffle_before_sampling in addition to sample_size. When
shuffle_before_sampling is True, the reader will shuffle the entire dataset before sampling, but
this requires fetching the entire dataset. If your dataset is massive and you only want to load the
first sample_size rows, set shuffle_before_sampling to False. When you have many shards and
shuffle_before_sampling is False, the reader will try to sample a few rows from each shard, to
avoid any shard skew.

I tested this by making 4 shards, each 180MB = 720MB dataset and uploading to s3://lilac-public-data/test-*.parquet. Loading with sample_size=1000 and shuffle_before_sampling=True takes ~2min since it has to download the entire dataset.
However with shuffle_before_sampling=False, it takes 4secs since it uses range requests to partially read the files.

Fixes #770
Fixes #779

[brilee]

When I call sample(), I think I expect to get a uniform random sampling, whereas this seems to be heavily implementation dependent - smaller shards will be overrepresented; you'll get multiple samples from one shard before moving on to the next, etc.

Still, this is probably a good enough implementation to move ahead for now.

[nsthorat]

wooop nice work!