sret unnecessarily blocks tail calls on X86

```llvm
%"string" = type { i64, i64, ptr }

define void @string_ctor(ptr sret(%"string") %agg.result) {
tail call void @func(ptr %agg.result)
ret void
}

declare dso_local void @func(ptr)
```
Currently causes a `call func` on X86, even though it should be perfectly safe to tail-call AFAICT. For some reason the `sret` breaks the tail call.

This has been reduced from libc++'s `string` implementation.

Start by reproducing the provided LLVM IR on X86 and inspect the tail-call lowering path for functions with an sret parameter. Done means the generated X86 code uses a tail call to func while preserving correct sret behavior, with a regression test covering the reduced case.

Written by the indexing model from the issue text.