numpy: Overloads between scalars and arrays can produce confusing results

[EricCousineau-TRI]

Relates numpy/numpy#10404 - can transplant the main info here if need be

Given how NumPy permits converting arrays of size 1 to scalars (regardless of dimension), this can create confusing interplay between overloads of an array and a scalar numeric value, especially if the array type is meant to be implicitly convertible from int or float.

Reproduction code:

sm.def("overload_scalar", [](py::array_t<float>) { return "Vector"; });
sm.def("overload_scalar", [](int) { return "Int"; });

def test_overload_scalar():
    assert m.overload_scalar(0) == "Int"
    assert m.overload_scalar([0]) == "Vector"
    assert m.overload_scalar(np.array([0.])) == "Int"  # What???

As a workaround, a user could define some wrapper type, like scalar_only<T>, and an accompanying type_converter<scalar_only<T>>, which explicitly rejects any containers that are iterable.

Regarding how to solve this, I unfortunately do not have any good ideas (though it would be nice if/when numpy/numpy#10615 lands).

[EricCousineau-TRI]

Added something like the above workaround to the current PR: 5e75f0b

EDIT: Note that this can also be workaround by having required kwargs (is this easily supported by pybind11?), or by writing the dispatch manually by writing your function on py::object and doing your own overload (losing out on the nice error messages, though).

[eacousineau]

Ran into some form of this (I think) when looking at #1785 / #1776.