load_table_from_dataframe -- support alternative serialization formats

[tswast]

Currently, pandas-gbq doesn't use load_table_from_dataframe because it is using CSV for serialization, which has better support for TIME, DATE, DATETIME than Parquet. See: #56, #382

Struct / Array support could also be improved (though this is partially due to problems in the pyarrow library). #19

I propose supporting the CSV serialization type.

[tswast]

To go from DataFrame to CSV, we can use the same tempfile logic as we have for parquet, but instead use to_csv + a few parameters to ensure proper data types:

https://github.com/pydata/pandas-gbq/blob/ac2d2fe4ac0025109f8df3723e3f03a337face94/pandas_gbq/load.py#L16-L30

[tswast]

This would also require adding CSV to the allowed serialization formats here:

python-bigquery/google/cloud/bigquery/client.py

Lines 2201 to 2206 in 809e4a2

	if job_config.source_format != job.SourceFormat.PARQUET:
	raise ValueError(
	"Got unexpected source_format: '{}'. Currently, only PARQUET is supported".format(
	job_config.source_format
	)
	)

[cguardia]

@tswast I have the initial work for this in #399. Looking for comments about which are the best parameters to use in the to_csv call, and whether this is going in the right direction.