Reject duplicate lifetime parameters with E0403

[u7k4rs6]

gccrs currently accepts struct Foo<'a, 'a>, for<'a, 'a>, etc. rustc rejects these with E0403, so this adds the check.

Two commits: first adds visit_generic_params / visit_for_lifetimes to DefaultASTVisitor (pure refactor, no behavior change), then overrides both in TopLevel to actually report E0403.

Went with TopLevel since it's the pass that collects definitions and enforces uniqueness (per Arthur on Zulip), while Late is for resolving uses.

Check is scoped to one binder, so an inner for<'a> reusing an outer 'a is untouched, that's shadowing (E0496), different issue.

TopLevel visits each item 3x, so I track reported params in NameResolutionContext to avoid tripling the error. This 3x-visit thing is an existing bug on master too, unrelated to this PR:

$ gccrs probe.rs      # extern crate does_not_exist;
error: unknown crate 'does_not_exist'
error: unknown crate 'does_not_exist'
error: unknown crate 'does_not_exist'

Output matches rustc:

error: the name ''a' is already used for a generic parameter in this item's generic parameters [E0403]
struct Foo<'a, 'a>(&'a u8);
           ~~  ^~

Tests: 11646 pass / 0 fail (+7 new, covering struct/fn/trait/type alias/impl/HRTB plus the shadowing and distinct-lifetime cases that should stay silent).

[CohenArthur]

I think this is a good start! The changes to the visitor are great, just need to make sure everything aligns together properly :) good work!!

[CohenArthur]

Do we not define lifetimes in the Labels namespace? Why can't we do something like insert_or_error_out(...) like in the rest of TopLevel?

[CohenArthur]

Likewise here if this is a definition spot

[u7k4rs6]

@CohenArthur We don't currently insert lifetime params into Labels, and Labels is shadowable for loop labels, so insert_or_error_out wouldn't catch duplicate lifetime params.

I'm happy to make lifetime params proper non-shadowable definitions instead. That would need a small change to the Labels insertion path and E0403 handling. Would you prefer that in this PR?

[u7k4rs6]

@CohenArthur took your suggestion lifetimes are definitions in the Labels namespace now rather than a separate check.

Two things made that work. ForeverStack<Namespace::Labels>::insert is specialised to shadowable so loop labels can nest ('a: loop { 'a: loop {} }), which meant insert_or_error_out would silently accept a duplicate param. Added an insert_non_shadowable alongside the existing insert_shadowable and used it only for params, so loop labels keep their current behaviour.

The second is scoping: with every binder sharing one rib, struct A<'a>(&'a u8) where for<'a> &'a u8: Tr; started erroring, but that's shadowing (E0496), not a duplicate. Each for<...> binder now gets its own Rib::Kind::Generics rib, so inner binders shadow instead of colliding.

Net effect is smaller than before check_duplicate_lifetimes and the reported_duplicate_lifetimes set in NameResolutionContext are both gone, since the fixed-point dedup falls out of the rib. make check-rust is clean (5529 passes, 0 failures) and CI is green.