Missed optimization: sub-optimal std::string::size codegen

std::string::size implementation:
https://github.com/llvm/llvm-project/blob/d2a097d6a36e28e6b2d8b2fcdb7c0fafe5e08e6a/libcxx/include/string#L1289-L1291

Currently clang generates:
```
movzx eax, byte ptr [rdi]
test al, 1
je .LBB1_1
mov rax, qword ptr [rdi + 8]
ret
.LBB1_1:
shr eax
ret
```
https://godbolt.org/z/Eq4x4rsY9

I think this could be (and it would be more efficient: less instructions/code, fewer instructions executed in the "long" case):
```
movzx eax, byte ptr [rdi]
shr eax
jnc .LBB1_1
mov rax, qword ptr [rdi + 8]
.LBB1_1:
ret
```

Start with the std::string::size implementation in libcxx/include/string at the linked lines, then inspect the Godbolt example and compare Clang's generated assembly. Determine whether the proposed flag-based control flow is valid and consistently improves the long-string case. Done means the missed optimization is addressed with appropriate LLVM project validation.

Written by the indexing model from the issue text.