BigQuery: Support user-overridable dtypes in to_dataframe method.

[tswast]

With pandas 0.24.0 (unreleased), a new dtype is available to support nullable integer columns. http://pandas-docs.github.io/pandas-docs-travis/integer_na.html#integer-na The default behavior is to convert to float, but this can result in data loss (#6177). This new dtype extension avoids that.

I propose we allow the user to provide a map from column names to dtypes for any columns for which they'd like to override the default behavior. This argument could be called dtype_overrides. This would also be useful for other extension dtypes in the future, such as for GEOGRAPHY columns.

See googleapis/python-bigquery-pandas#242 for additional discussion.

Alternatives

• Make the new dtype for nullable integer the default for integer columns.
  • Con: Not compatible with older versions of pandas.
  • Con: Inconsistent with pandas's default behavior.

[tswast]

I experimented with this in master...tswast:b122674716-bqstorage-types for the BigQuery Storage API. Similar work is needed for the BigQuery API.

I think the dictionary of column names to dtypes works well. I don't see any problems with using pandas Series constructor for type casting.

Should it be an error if a dtype was supplied by the column isn't actually in the DataFrame? I think we might be able to parse the avro_schema to see what columns are available ahead of time.