API: combine casting behavior with extension types

[mzeitlin11]

In combine, current behavior tries to maintain extension types by casting back if possible through maybe_cast_pointwise_result. This can give some value dependent behavior with a (contrived) example like the following:

In [3]: ser = pd.Series([True, True], dtype="boolean")

In [4]: ser.combine(2, lambda x, y: int(x) * y)
Out[4]:
0    2
1    2
dtype: int64

In [5]: ser = pd.Series([False, False], dtype="boolean")

In [6]: ser.combine(2, lambda x, y: int(x) * y)
Out[6]:
0    False
1    False
dtype: boolean

Not sure if that is desired behavior on its own, but it presents an issue for fixing #42137. On current master object boolean data raises in IntegerArray._from_sequence, so combine fails to cast back:

In [2]: ser = pd.Series([1, pd.NA], dtype="Int64")

In [3]: ser.combine(2, lambda x, y: x < y)
Out[3]:
0    True
1    <NA>
dtype: object

However, fixing #42137 will allow casting back, changing this behavior to

In [5]: ser = pd.Series([1, pd.NA], dtype="Int64")

In [6]: ser.combine(2, lambda x, y: x < y)
Out[6]:
0       1
1    <NA>
dtype: Int64

which looks unexpected (and fails pandas/tests/extension/test_integer.py::TestMethods::test_combine_le section of tests).

[jbrockmendel]

I think the answer is "yes, this is expected"