Win64: Fix extern(C++) ABI wrt. passing non-POD structs by value

[kinke]

Return types have been checked for POD-ness already, but parameters haven't.

[kinke]

The dmd-testsuite tests work for Win32 and Win64, so both ABI (LL ref vs. value) and destruction rules (MSVC++ lets the callee destruct the param, just like D) match across D and C++, which is pretty nice. Maybe there's a slight chance that 32-bit POSIX passes the tests too, I haven't checked yet.

There's still at least one remaining special case not covered yet, and that's MSVC++ apparently treating struct S { int a; ~S(); }; as POD (any ctor will make it a non-POD, AFAIK another MSVC special case, but a dtor alone doesn't), while the D frontend doesn't.

[kinke]

Maybe there's a slight chance that 32-bit POSIX passes the tests too

Nope, 32-bit Linux fails with a segfault; like for x86_64, a ref is passed, and the caller destructs it.

[kinke]

Erm, turns out this only fixes the case where there's a deleted copy ctor and a move ctor, where 64-bit MSVC++ really passes a ref. With copy ctor, the argument is passed by value in a register, but the return value is still returned via hidden sret (!!), so the existing code and the inconsistent POD-ness check for return/param types is actually correct. A deleted C++ copy ctor apparently enforces by-ref passing though...

[kinke]

I reworked this, especially the tests. They don't handle all permutations and corner cases of course, but the resulting fix should make byval-interop with C++ on Windows/MSVC work in most cases now.

[kinke]

[Slightly reworked once more.]