Rollup merge of #161467 - Dnreikronos:wfcheck/unused_param_name_discarding_item, r=fmease

wfcheck: name the item that discards an unused type parameter

fixes rust-lang/rust#161365

label was "`T` is named here, but is likely unused in the containing type". the "likely" is the part that bugs me. we can just go look, so now it does:

```
LL |     bar: Bar<T>,
   |              - `T` is named here, but `Bar` does not use it
```

while collecting the usage spans it holds on to the outermost thing that throws the argument away. adts are easy, bivariant param in variances_of. aliases have no variances so you expand them and see what's left, and my first version of that only looked for the param in the rhs, which is wrong. `type Alias<T> = DropFirst<T, u8>` mentions T and still drops it. so it walks the expansion skipping bivariant slots. (nested args blame the outer one, `Bar<Vec<T>>` points at Bar.) if it can't work out who's responsible it falls back to "this does not constrain `T`".

warning for whoever touches this next: type_of doesn't expand free aliases and you can't chase them yourself. `type Recur<T> = Recur<(T,)>` never repeats its args so there's nothing to stop on, stack goes. expand_free_alias_tys is the one with the limit. variance-overflow.rs has a comment at the top saying exactly this and i wrote it wrong anyway, so i stuck a case in there for the diagnostic path.

tests in variance-unused-type-param.rs, new file for alias chains.

there's a separate bug in the "only used recursively" branch right next to this one that i left alone, wrote up why in a comment.