BUG: fix timedelta floordiv with scalar float (correction of #44466)

[jorisvandenbossche]

Follow-up on #44466 (comment)

I still need to add a test that would actually catch this

[jorisvandenbossche]

Actually, I don't directly find a way to edit the test that it would have catched this. It seems that in practice timedelta division already acts similar as a floor division:

In [16]: np.array([1, 2, 3, 4, 5], dtype="m8[ns]") / 3
Out[16]: array([0, 0, 1, 1, 1], dtype='timedelta64[ns]')

In [17]: np.array([1, 2, 3, 4, 5], dtype="m8[ns]") // 3
Out[17]: array([0, 0, 1, 1, 1], dtype='timedelta64[ns]')

In [18]: np.array([1, 2, 3, 4, 5], dtype="m8[ns]").astype(float) / 3
Out[18]: array([0.33333333, 0.66666667, 1.        , 1.33333333, 1.66666667])

(so the integers for timedelta divison result already seem to be obtained with a floor round)

[jorisvandenbossche]

Actually with a round towards zero (so np.fix and not np.floor), but that is still done for both division and floor division.

[jorisvandenbossche]

To conclude: this can be merged as is?

(the one failure is a CSV compression failure, so I assume unrelated)

[jbrockmendel]

later today ill look at the scalar timedelta tests to see if past-me came up with anything.

[jbrockmendel]

BTW @jorisvandenbossche can you take a look at #43840

[jbrockmendel]

Looking at the scalar Timedelta.__floordiv__ i dont see anywhere where this makes a difference.

LGTM

[jorisvandenbossche]

OK, merging this then