Tweak date/datetime inference, remove dependency on non-core Numpy/Pandas API

[sc1f]

This PR fixes the inference of dates and datetimes for pandas DataFrames, specifically for DatetimeIndex of various frequencies. Instead of inferring the data type based on the datetime64 unit, we use the index's datetime frequency to decide whether to type the column as a date or a datetime.

Other fixes:

• remove fillna from our dataframe processing flow
• remove dependencies on masked_invalid and non top-level numpy/pandas API
• enforce all column names as strings, and throw an error otherwise
• add new tests + util.make_dataframe and util.make_series as test utilities

Reviewed offline