REGR: fix regression in scalar setitem with setting a length-1 array-like

[jorisvandenbossche]

• closes REGR: ValueError: setting an array element with a sequence. when assigning an array.array using .loc #46268
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

While testing the various related cases, I noticed that we don't have very consistent behaviour always, see the TODO notes in the tests.
In pandas 1.3, we always unpacked an array-like when setting, also for object dtype. This means that for object dtype, it unpacked it for len-1, and raised an error for longer arrays:

In [31]: df = pd.DataFrame({"A": [1, 2, 3], "B": np.array([1, 'a', 'b'], dtype=object)})

In [32]: df.loc[0, "B"] = np.array([0])

In [33]: df
Out[33]: 
   A  B
0  1  0
1  2  a
2  3  b

In [34]: df.loc[0, "B"] = np.array([0, 1])
...
ValueError: Must have equal len keys and value when setting with an iterable

With the change in #42780, in pandas 1.4 we started to preserve the nested object in case of len-1, but still raising for longer arrays (with object dtype column):

In [24]: df = pd.DataFrame({"A": [1, 2, 3], "B": np.array([1, 'a', 'b'], dtype=object)})

In [25]: df.loc[0, "B"] = np.array([0])

In [26]: df
Out[26]: 
   A    B
0  1  [0]   # <-----
1  2    a
2  3    b

In [27]: df.loc[1, "B"] = np.array([0, 1])
...
ValueError: Must have equal len keys and value when setting with an iterable

This is a bit inconsistent I would say, and I think we should probably long term allow setting nested data of any length in case of object dtype column (the first TODO note).

Because of the above change (no longer unpacking for len-1 array), this started to error for numeric columns (i.e. the actual regression reported in #46268).

What I did in this PR is limiting the change of #42780 to only object dtype, so we keep the new behaviour of preserving the nested object when setting in case object dtype, while fixing the regression for numeric data.
(strictly speaking, we could also say that it is a regression for the object dtype case, but there I think it was actually an improvement in behaviour, and since this was already changed in 1.3.5, and nobody reported it, it seems better to keep that now)

In addition, while testing the same for Series, I noticed that for Series setitem, we only do this unpacking of len-1 array-like values for actual numpy arrays, and not for setting array-likes in general (the second TODO note).

TODO: open new issues to keep track of those TODOs and update those comments

[phofl]

Some comments

[phofl]

I think this sound good

[phofl]

If we do the above, this is correct I think

[phofl]

Yes, I think this should unpack for everything. The DataFrame case already does

[jorisvandenbossche]

@phofl just to be clear: those TODO comments I added are interesting points (and I should open new issues for those), but fixing those are out of scope of this focused regression fix for 1.4.4 (your feedback on those questions was very welcome! just to set expectations that I don't plan to "fix" those todos here ;)). See also the longer explanation in the top post.

[phofl]

Yep understood this, no worries. Just wanted to coment in code to avoid misunderstandings with the references

[phofl]

I think this is good to merge

[mroeschke]

Thanks @jorisvandenbossche (failures unrelated)