Rollup merge of #135766 - lcnr:candidate-assembly-3, r=compiler-errors

handle global trait bounds defining assoc types

This also fixes the compare-mode for
- tests/ui/coherence/coherent-due-to-fulfill.rs
- tests/ui/codegen/mono-impossible-2.rs
- tests/ui/trivial-bounds/trivial-bounds-inconsistent-projection.rs
- tests/ui/nll/issue-61320-normalize.rs

I first considered the alternative to always prefer where-bounds during normalization, regardless of how the trait goal has been proven by changing `fn merge_candidates` instead. https://github.com/rust-lang/rust/blob/ecda83b30f0f68cf5692855dddc0bc38ee8863fc/compiler/rustc_next_trait_solver/src/solve/assembly/mod.rs#L785

This approach is more restrictive than behavior of the old solver to avoid mismatches between trait and normalization goals. This may be breaking in case the where-bound adds unnecessary region constraints and we currently don't ever try to normalize an associated type. I would like to detect these cases and change the approach to exactly match the old solver if required. I want to minimize cases where attempting to normalize in more places causes code to break.

r? `@compiler-errors`

Author: matthiaskrgr