Implement P2438R2 string::substr() &&

[frederick-vs-ja]

And make the old and new substr use default constructed allocators.

(And consistently refer to WG21-N4910 in <xstring>.)

Fixes #2928.
Fixes #3022.

As implementation details, the rvalue substring constructors reuse the source storage if and only if

• the provided allocator and the allocator of the source compare equal, and
• the constructed string doesn't fit into SSO buffer.

I have no idea how to portably test the probably intended behavior of newly added overloads...

[frederick-vs-ja]

It seems that EDG is buggy here:

• EDG's __builtin_memcmp seemly doesn't work well with constexpr-dynamically-allocated storage,
• EDG seemly mishandled overload resolution between T::operator const U& and T::operator U&&.

I'll skip relative cases for EDG.

[StephanTLavavej]

Thanks, this looks great! I'll validate and push changes for the minor issues and nitpicks I found.

[StephanTLavavej]

I'm speculatively mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[CaseyCarter]

Going to approve because I don't want to reset testing, but I'll address these comments in a followup.

[StephanTLavavej]

I've resolved a trivial adjacent-add conflict in tests/std/test.lst.

[StephanTLavavej]

Thanks for implementing this feature in one of the STL's most-used data structures! 😻 ✅ 🧶