<syncstream>: implement basic_syncbuf and basic_osyncstream

[MichaelRizkalla]

This PR will implement P0053R7 and P0753R2. should close #3.

It's still on going work, current progress tracking:

• Implement basic_syncbuf
• Implement basic_osyncstream
• Implement osyncstream manipulators
• Implement tests for basic_syncbuf
• Implement tests for basic_osyncstream
• Implement tests for osyncstream manipulators

[MichaelRizkalla]

I started this review to have all questions that I think need a maintainer's input.

[miscco]

There are some concerns about fancy pointers and allocators, but besides that mostly nits.

Great progress 👍

[MichaelRizkalla]

I started sorting out the tests for this PR.
I do not really understand why P1502R1_standard_library_header_units is failing now, so I'd like your input.

[StephanTLavavej]

Thanks for fixing the fancy pointer and allocator issues! I've pushed two commits, adding explicit to the ID class's constructor (it's test code, but I always recommend being cautious about implicitly converting constructors), and adding the necessary stl/msbuild changes. I've mirrored this to internal MSVC-PR-303125 along with the other necessary internal changes. 😸 (This is mirrored all by itself, not chained with other GitHub PRs, due to the magnitude of the changes and the importance of the feature.)

Please note: Further changes are still possible (e.g. if you notice something, or if final review leads to more feedback that should be addressed before merging), but they must be replicated to the mirror PR in order to avoid codebase divergence.

[StephanTLavavej]

I've pushed a comment change to an internal stl/msbuild file (technically unrelated to this PR, but right above a modification, and @CaseyCarter noticed that the comment was confusing).

[StephanTLavavej]

I've pushed and mirrored a change to use _Init_locks, which resolves the Standard Library Header Units crash (so we can restore that part of the test). There's a Static Initialization Order Fiasco, which was revealed but not inherently caused by modules, where the syncstream map requires the STL's locks to be initialized (because for IDL=2, we need to take the debug lock in map's destructor to orphan all of its iterators). If the locks aren't guaranteed to be initialized before map construction, then they aren't guaranteed to be destroyed after map destruction, and we're doomed.

The fix is to use the _Init_locks code pattern that's used elsewhere in the STL:

STL/stl/src/cout.cpp

Lines 9 to 11 in 8f79acf

	#pragma warning(disable : 4074)
	#pragma init_seg(compiler)
	static std::_Init_locks initlocks;

This guarantees initialization order within the syncstream.cpp translation unit. (I had seen this pattern before, but forgot about it since I've never added occurrences in my 14 years working on the STL - first time for everything.)

I have copied the exact spelling of the _Init_locks pattern for searchability, even though we should probably say _STD for consistency (that should be a global cleanup). The non-_Ugly identifier initlocks is fine because this is static (again, we could clean it up, but that would be a separate global cleanup).

[CaseyCarter]

Just a couple of noexcept questions, looks good otherwise.

[StephanTLavavej]

Thanks for implementing this new stream! 🎉 🚀 😺 This will ship in VS 2019 16.10 Preview 2.