Intrinsics should allow a non-const fallback body if hooked for CTFE

[tgross35]

Currently if an intrinsic is marked const, its body needs to be const as well. This makes sense when the body is providing the comptime implementation, but is limiting in cases where CTFE hooks the function so the fallback would only ever be called in non-const contexts.

Noticed in #150946

[RalfJung]

Cc @rust-lang/wg-const-eval

I think the hardest part here is to decide about the syntax to distinguish "const intrinsic with const fallback body" and "const intrinsic with non-const fallback body".

[tgross35]

I think the hardest part here is to decide about the syntax to distinguish "const intrinsic with const fallback body" and "const intrinsic with non-const fallback body".

Perhaps they should use the same syntax but a #[const_intrinsic_must_hook] attribute opts the second case out of a const body? I was going to suggest removing the const keyword in the for "const intrinsic with non-const fallback body", but then decided that making the signature appear correct to callers and in docs (given it can be called in const) is more important than matching up with whatever is going on in the body, since that's already pretty magic.

[RalfJung]

Yeah that sounds plausible.