Series.combine_first() misbehaves with Timestamp data

[asadovsky]

Best illustrated by example:

In [957]: s0 = pd.Series(["2010", np.NaN])

In [958]: s1 = pd.Series([np.NaN, "2011"])

In [959]: s0.combine_first(s1)
Out[959]:
0 2010
1 2011

In [960]: s0 = pd.to_datetime(pd.Series(["2010", np.NaN]))

In [961]: s1 = pd.to_datetime(pd.Series([np.NaN, "2011"]))

In [962]: s0.combine_first(s1)
Out[962]:
0 2221-02-23 04:49:47.750490112
1 2051-05-17 05:40:40.331386880

[wesm]

marked as bug, thanks

[changhiskhan]

added _where_compat in common.py that converts to i8 or python datetime first if necessary.