Boolean column with NAs becomes obj64 after roundtrip to pandas

[st-pasha]

If a boolean column with NAs is converted to pandas and then back to a datatable Frame, the end result is an obj64 column:

>>> import datatable as dt
>>> dt.Frame(dt.Frame([True, False, None]).to_pandas()).stypes
(stype.obj64,)
>>> dt.Frame(dt.Frame([True, False, None]).to_pandas())
       C0
--  -----
 0   True
 1  False
 2    nan

[3 rows x 1 column]

The problem here is that pandas does not support boolean columns with NAs, so it converts the column to object, and we retain this type when converting back. However, this is problematic because obj64 columns have very limited support in datatable: they cannot be sorted, saved to jay or csv, cannot compute stats, cannot participate in arithmetics, etc.

Note that integer columns have a similar problem: after round-trip to pandas they become float64. This, however, is not as bad because this float64 column is numerically equal to the original, and will produce similar results for most operations.