TST: Add astype_nansafe datetime tests

[dsaxton]

Adds tests for astype_nansafe in situations where we can't perform a datetime casting because of a precision mismatch:

arr                                                                             
# array(['2018-01-01'], dtype='datetime64[D]')

arr.view(np.int32)                                                              
# array([17532,     0], dtype=int32)

arr.view(np.float32)                                                            
# array([2.4568e-41, 0.0000e+00], dtype=float32)

Related to #28492

[WillAyd]

Could this get used elsewhere? If so might be good to make a shared fixture for numeric_dtype or something to the effect

[dsaxton]

This particular setup might be a little specific in that all the to_types have lesser precision than the from_types, but did you have something in mind?

[jreback]

this is ok

[jreback]

this is ok

[jreback]

can you add a test for the object path (for dtype=) when its a datetime/timedelta

[dsaxton]

What would the expected output be for that case? I would think it'd agree with astype from numpy when we don't have any NA values but the result is a bit odd:

[ins] In [3]: arr = np.array([np.datetime64("2018")])                                          

[ins] In [4]: arr.astype("object")                                                             
Out[4]: array([datetime.date(2018, 1, 1)], dtype=object)

[ins] In [5]: astype_nansafe(arr, dtype="object")                                              
Out[5]: array([datetime.datetime(1970, 1, 1, 0, 0)], dtype=object)

[jreback]

how is this PR distinct from #28492

[dsaxton]

how is this PR distinct from #28492

I was thinking it was worth treating separately since that PR deals with NaT specifically whereas this was testing other paths (non-missing values that nonetheless can't be cast) without any logic changes

[jreback]

i would just close this and merge #28492