BUG/TST: Fix #7259, do not "force scalar" for already scalar str/bytes

[gerritholl]

Fix for #7259. Since commit f8f753b, fill_value was forced scalar by
indexing with [()]. This works for most numpy scalars, but not for U
(str) and S (bytes). This commit falls back to the pre-f8f753b version
when forcing to scalar fails for fill_value of dtypes with kind U and S.
Also add a test to see that this works in practice.

[ahaldane]

Seems good, although a deeper fix would be to make scalar strings allow indexing with an empty tuple.

@gerritholl , are you interested in looking at a deeper fix, or would you rather commit this PR now and open a new issue to fix it properly later? I'm fine either way. In the latter case we should open an issue linking to this one, and add a comment in the code "(this can be removed once #XXXX is fixed)".

[ahaldane]

Here are some beginning investigations about how the deeper issue (tuple indexing of strings) can be fixed. I can't totally see the solution, and it's harder, but I think it involves the following in scalartypes.c.src:

A. Define a function stringtype_subscript (analagous to gen_arrtype_subscript), see below

static PyObject *
stringtype_subscript(PyObject *self, PyObject *key) {

B. Define (analagous to gentype_as_mapping)

static PyMappingMethods stringtype_as_mapping = {
    NULL,
    (binaryfunc)stringtype_subscript,
    NULL
};

C. in initialize_numeric_types do

PyStringArrType_Type.as_mapping = &stringtype_as_mapping;
PyUnicodeArrType_Type.as_mapping = &stringtype_as_mapping;

The part I don't see yet is what exactly to put in to stringtype_subscript. Probably it should check for an empty tuple and return self if so, otherwise call "super" on the string/unicode getitem/setitem. But I'm not sure how the super call works. I also haven't figured out how/where the numpy scalar-type inheritance is set up.

[gerritholl]

I agree that this might be a better fix. Unfortunately, I have very little experience writing C and no experience at all writing C/Python, so I don't consider myself the right person to implement the change such as you propose.

[ahaldane]

All right, I'd like to merge then, after a possible tweak.

Could we change it to

if isinstance(self._fill_value, ndarray):
    return self._fill_value[()]
return self._fill_value

instead of the try/catch, because it maybe makes clear we are accounting for the ndarray vs scalar cases, if you agree?

[charris]

@gerritholl Could you make the suggested fix?

[ahaldane]

LGTM, merging. Thanks @gerritholl.

(By the way I think this is the correct fix even if #7267 is solved, since #7267 cannot fix the case of object arrays)