Update tests to new style

[jagerman]

This updates the test suite tests to the new style proposed and started in #898. This is still in progress. (I'm adding commits one-by-one for now, but intend to merge them all into one before merging onto master).

(Also note that test_exceptions and test_multiple_inheritance won't be included here—they are covered by PRs #954 and #960, respectively.)

[jagerman]

This PR broke the current py::overload_cast with ICEs under both MSVC 2015 and under clang 3.8 on debian stretch.

I spent some time fighting with MSVC to get it to work, and finally settled on e4fb05a, which manages to simplify it, fix the ICEs on MSVC and clang, and makes it C++11-compatible. I came on the really weird trick to making it work accidentally: dropping the defaulted std::false_type argument makes it work. For some bizarre reason it made MSVC choke on the deductions. But having two functions here--one with the std::false_type argument and one without--makes MSVC perfectly happy.

@dean0x7d - do we need the std::false_type argument at all? We could just drop the second function that takes it, but I'm assuming you had a reason for adding the argument in the first place.

[dean0x7d]

The solution in e4fb05a will not work for overload sets which have a variable number of arguments, e.g. void foo() and void foo(int). That's the main reason why the two stage template parameters are needed (class template parameters + static member template parameters). See the test.cpp file in https://github.com/dean0x7d/overload_cast for the relevant tests.

The std::false_type argument is there to allow programmatic selection of const vs non-const overloads, based on some type trait. E.g. overload_cast<int>(&W::foo, std::is_same<T, U>{}). It's not needed for manual casting, but it's pretty nice for metaprogramming.

[jagerman]

Rats. I think we may have to switch off overload_cast for MSVC 2015 (at least in the test suite). There's nothing inherently unusual about the use in the test code; the slightly different context just seems to trigger an ICE.

[jagerman]

I think I've found a fix (for the original overload_cast implementation) for MSVC 2015: it seems adding an explicit constexpr cast_overload_impl() {} makes it happy. (But changing the {} to = default; makes the ICE happen again).