Combining place projections in MIR makes codegen forget that NonNull is nonnull

[saethlin]

In #105808 I am adding a hack to prevent combining place projections where one of assignments that would be optimized away contains a niche. This is because optimizing Vec::as_ptr to a single MIR statement that just loads the underlying *const T makes us lose the nonnull LLVM attribute in codegen.

(I am opening this issue early so I can link to it in the PR)

[scottmcm]

The reminds me of why I dislike the "it's just a field" approach in the first place (and thus I dislike rust-lang/rfcs#3334), and reminds me of similar problems that repr(simd) hit with looking at fields (see https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/splat.20no.20longer.20compiles.20for.20release.20builds/near/351948961).

So maybe we should use the same approach that repr(simd) is moving towards: just disallow place projections to the inner field entirely for such types. That makes this issue irrelevant -- as the project isn't there to be combined -- and the niche support ends up in the transmutes and such instead, where it sees the different types properly instead of just projections.