Support INTERVAL data type in DB-API, arrow, and pandas connectors

[tswast]

Follow-up to #826 since pandas and Arrow do not yet have a structured type that aligns with INTERVAL. The existing Timedelta support would work for INTERVALS with only a time component, but it is not calendar aware, so supporting year, month, and day intervals would require some mapping to timedelta, which is not ideal.

Why is a new data type needed?

• YEAR: Leap years are a thing. Not every year is 365 days long.
• MONTH: Not every month is the same length.
• DAY: Daylight savings is a thing. Not every day is 24 hours long.

Note: DB-API support is included here because it uses the BigQuery Storage API, where we use the Arrow wire format.

TODO:

• Auto-detect data type in DB-API query parameters
  • Might be possible to do this before reading INTERVAL columns is supported.
• Row data is converted to relevant type in DB-API
• Row data is converted to relevant type in to_dataframe
  • Might need to be object, since timedelta64 doesn't have years/months.
• Check if to_arrow type is expected datatype
• Convert data type in insert_rows_from_dataframe
• Convert data type in load_rows_from_dataframe (CSV)
• Convert data type in load_rows_from_dataframe (Parquet)

[tswast]

Update: Maybe pandas does have the right data type? https://pandas.pydata.org/docs/reference/api/pandas.tseries.offsets.DateOffset.html

[tswast]

Note: There is currently a bug with the interval type in pyarrow. https://issues.apache.org/jira/projects/ARROW/issues/ARROW-15783

Sounds like we need to call pa.array([1]) to initialize some pandas code before attempting to convert an interval column to pandas.