BUG: Fixed IntervalArray[int].shift

[TomAugspurger]

Closes #31495

In ExtensionArray.shift, we have the note

        # Note: this implementation assumes that `self.dtype.na_value` can be
        # stored in an instance of your ExtensionArray with `self.dtype`.

I wonder, should we have a method / property like

@property
def _can_hold_na_value(self):
    return True

And for IntervalArray, we would do something like

@property
def _can_hold_na_value(self):
    return is_float_dtype(self.dtype.subtype)

That would let us deduplicate things, since the call to _from_sequence would know to not pass dtype=self.dtype and trigger a re-inference.

[TomAugspurger]

cc @jschendel for thoughts on that.

[TomAugspurger]

This is still failing for interval[datetime64], since we try to concat a float (all NaN) and datetime. Will poke around a bit.

[TomAugspurger]

The datetime64 failure is present on master. I've opened #31504 for it.

[jbrockmendel]

LGTM, agreed it would be nice to find a way to make this work in the general case.

[TomAugspurger]

All green now. Sorry about the string of CI failures earlier.

[jschendel]

lgtm, thanks