REF: Block._astype defer to astype_nansafe in more cases

[jbrockmendel]

Makes astype_nansafe for (td64|dt64) -> (object|str|string) match DTA/TDA/Series behavior.

Medium-term (weeks) the goal is to get rid of Block._astype altogether and just use astype_nansafe, which among other things will be helpful for ArrayManager.

This changes Series[dt64].astype("string") behavior in a way that causes a new xfail in test_astype_roundtrip, but as discussed in #36153 that test is already wrong for other reasons.

This also has a side-effect of changing Series(dt64, dtype="Sparse[object]") behavior, discussed in #38508 as possibly not-desirable.

[jreback]

This also has a side-effect of changing Series(dt64, dtype="Sparse[object]") behavior, discussed in #38508 as possibly not-desirable.

where is this test case changed?

[jbrockmendel]

where is this test case changed?

we dont have a test for this; #38508 introduced new ones

[simonjayhawkins]

This changes Series[dt64].astype("string") behavior in a way that causes a new xfail in test_astype_roundtrip, but as discussed in #36153 that test is already wrong for other reasons.

I'm not sure about the new behaviour, I think this should at least have a release note if not reverted.

old behaviour

>>> pd.__version__
'1.2.4'
>>> 
>>> tdi = pd.timedelta_range("1 Day", periods=3)
>>> ser = pd.Series(tdi)
>>> ser
0   1 days
1   2 days
2   3 days
dtype: timedelta64[ns]
>>> ser.astype("string")
0    1 days
1    2 days
2    3 days
dtype: string
>>> 
>>> dti = pd.date_range("2021", periods=3)
>>> ser = pd.Series(dti)
>>> ser
0   2021-01-01
1   2021-01-02
2   2021-01-03
dtype: datetime64[ns]
>>> ser.astype("string")
0    2021-01-01
1    2021-01-02
2    2021-01-03
dtype: string

new behaviour

>>> pd.__version__
'1.3.0.dev0+1567.g67c9385787'
>>> 
>>> tdi = pd.timedelta_range("1 Day", periods=3)
>>> ser = pd.Series(tdi)
>>> ser
0   1 days
1   2 days
2   3 days
dtype: timedelta64[ns]
>>> ser.astype("string")
0     86400000000000 nanoseconds
1    172800000000000 nanoseconds
2    259200000000000 nanoseconds
dtype: string
>>> 
>>> dti = pd.date_range("2021", periods=3)
>>> ser = pd.Series(dti)
>>> ser
0   2021-01-01
1   2021-01-02
2   2021-01-03
dtype: datetime64[ns]
>>> ser.astype("string")
0    2021-01-01T00:00:00.000000000
1    2021-01-02T00:00:00.000000000
2    2021-01-03T00:00:00.000000000
dtype: string
>>> 

[jreback]

@simonjayhawkins can u open a new issue - this is going to be very hard to revert - but i agree original behavior is correct so should fix

[simonjayhawkins]

#41409