Fix ordering for `default impl` check in the new solver

The second commit of #160605 moved the `default impl` check later, for
better performance, which introduced a regression. This commit moves the
check a little earlier, so it is after the `args_may_unify` call (thus
retaining the perf benefit) but before the `probe_trait_candidate`
(which has side-effects).

The check is now duplicated in three `GoalKind::consider_impl_candidate`
methods, which is unfortunate, but it fits in with the existing
duplicated code in those methods. And it means another copy of the
check (in `try_assemble_bounds_via_registered_opaques`) can be removed.

Fixes #160994.