name mangling and TypeId should not pass for aliases and generic params

[lcnr]

Name mangling currently succeeds even if the mangled type contains projections or opaque types. I am not sure whether TypeId has the same issue.

This can result in linking errors or potential UB (cc #97156) and we should guard against this by failing to mangle in this case. This change is non-trivial as we currently have name mangling tests which mangle generic modules. These tests should be changed.

cc @michaelwoerister @eddyb

[BoxyUwU]

Does -Zpolymorphization not require us to mangle generic params? I thought that we used generic params to represent unused substs

[lcnr]

yes, I also believe that longterm we should change polymorphization to instead use Bound variables. imo using ty::Param for polymorphization makes it too easy to break substitution (or hide a missing substitution)