DEPR: box argument in to_datetime and to_timedelta

[mroeschke]

Following @jreback #24415 (comment), we should deprecate the box argument in to_datetime and to_timedelta and only return pandas data structures. There already exists options to convert to numpy data structures via .values or to_datetime64/to_timedelta64

[shoyer]

We find the option box=False quite useful in xarray, to let us leverage pandas for converting inputs into NumPy's timedelta64 and datetime64 dtypes.

I would be a little sad to see it go away, though of course coercing into NumPy arrays is pretty orthogonal to the goals of pandas.

[jreback]

use .to_numpy() if you must

[jorisvandenbossche]

You cannot use to_numpy for scalars, as the example of @shoyer is in #24957 (comment)

(with 0.23.4)

In [58]: pd.to_timedelta('NaT', box=False)
Out[58]: numpy.timedelta64('NaT','ns')

In [59]: pd.to_timedelta('NaT', box=True)
Out[59]: NaT

In [60]: type(pd.to_timedelta('NaT', box=True))
Out[60]: pandas._libs.tslibs.nattype.NaTType

[jorisvandenbossche]

OK, but I see there is #24653 for adding to_numpy to scalars as well.

For the specific case above, however, the problem is that we don't have a NaT for timedelta, so you would still get np.datetime64('NaT', 'ns')