Make interpolate_2d handle datetime64 correctly

[jbrockmendel]

Broken off from #27626 because I decided that is poorly scoped.

[gfyoung]

Fix looks relatively straightforward, just need tests and whatsnew

[jbrockmendel]

Test added, unnecessary piece removed. No release note needed since no user-facing changes.

[TomAugspurger]

LGTM overall. I noticed the following while testing locally

In [1]: import pandas as pd

In [2]: dti = pd.date_range("2015-04-05", periods=5, tz="US/Central")
   ...: ser = pd.Series(dti)
   ...: ser.iloc[1:4] = pd.NaT
   ...: ser.astype('datetime64[ns]').to_frame().interpolate(method="pad", limit=2)
Out[2]:
                    0
0 2015-04-05 05:00:00
1                 NaT
2                 NaT
3                 NaT
4 2015-04-09 05:00:00

That happens on both master and this PR. Do you know if we have an issue for it? The important bits are

1. datetime64[ns]-dtype. datetime with TZ seems fine
2. DataFrame.interpolate. Series[datetime64[ns]].interpolate seems fine

[jbrockmendel]

datetime64[ns]-dtype. datetime with TZ seems fine
DataFrame.interpolate. Series[datetime64[ns]].interpolate seems fine

I'm not sure I understand your example. Is it relevant that you passed tz to date_range? You subsequently astype-d it away.

Best guess: DatetimeTZBlock.fillna dispatches to DatetimeArray.fillna, so doesn't go directly through here.

[jbrockmendel]

@jreback gentle ping. Between this and the quantile PR, we're about to be about to be able to rip out a bunch of _try_coerce_result code

[TomAugspurger]

Thoughts @jreback?

[jreback]

thanks