Support configuring types_mapper in read_gbq

[bnaul]

googleapis/python-bigquery#1529 and googleapis/python-bigquery#1547 have recently added arguments for overriding the default type conversions performed by record_batch.to_pandas(); this allows, for example, loading string data directly into dtype string[pyarrow] (which can be quite a bit more efficient) without doing any expensive conversions after the fact.

I think we could basically just copy the implementation from the above PRs, same kwarg names and everything. Anyone see any potential issues @jrbourbeau @j-bennet @ncclementi?

[j-bennet]

@bnaul You can already provide a custom mapper in read_dbq as part of arrow_options:

https://github.com/coiled/dask-snowflake/blob/42cb99e4e35aeb18f7ede95badba8f20a4113378/dask_snowflake/core.py#L215-L217

[bnaul]

Ha well this looks great, but unfortunately that's dask-snowflake and this is dask-bigquery 😅 but yeah that's basically what I had in mind! Do you think the same arrow_options approach makes sense here? As opposed to copying the many many kwargs that they expose in google.cloud.bigquery.QueryJob.to_dataframe().

[j-bennet]

:) :) :) scratch that, I'm contributing to dask-snowflake and dask-bigquery right now, and completely confused which one we're talking about.

Yes, I think arrow_kwargs would make sense in read_dbq. It would not take much, we would need to pass those kwargs through to the point where we call to_pandas on a pyarrow record batch:

dask-bigquery/dask_bigquery/core.py

Lines 45 to 48 in 876ddb8

	pyarrow.ipc.read_record_batch(
	pyarrow.py_buffer(message.arrow_record_batch.serialized_record_batch),
	schema,
	).to_pandas()