More robust type_eq_non_static implementation

[dtolnay]

According to #6 (comment) in reference to the type_id_of::<T> as usize-based implementation:

As for the function instantiation address taking thing: it simply looks broken/unsound - the only reason LLVM mergefunc would even consider them distinct is its implementation being too simple and nowhere near "graph isomorphism" (remember that we have unnamed_addr semantics for functions, which makes it valid to merge functions even if their addresses are observed).

I am being told by lcnr (also working on polymorphization) that this will break at some point the future (assuming we get to turn that on by default), so this is not that theoretical.

[dtolnay]

1.38.0 macos failure in CI is rust-lang/rust#105167, unrelated to this PR.

[sagebind]

Alright, CI should pass now if you pull in the latest master commits into your branch.

[sagebind]

Looks reasonable to me. The prior implementation always did make me uncomfortable.

[sagebind]

Thanks for working on this!