Revert performance-sensitive change in #82436

This change was done in #82436, as an "optimization". Unfortunately I
missed that this code is not always executed, because of the "continue"
in the conditional above it.

This commit should solve the perf regressions introduced by #82436 as I
think there isn't anything else that could affect runtime performance in
that PR. The `Pick` type grows only one word, which I doubt can cause up
to 8.8% increase in RSS in some of the benchmarks.

Author: osa1

compiler/rustc_typeck/src/check/method/probe.rs

@@ -742,16 +742,16 @@ impl<'a, 'tcx> ProbeContext<'a, 'tcx> {
 
         debug!("assemble_inherent_impl_probe {:?}", impl_def_id);
 
-        let (impl_ty, impl_substs) = self.impl_ty_and_substs(impl_def_id);
-        let impl_ty = impl_ty.subst(self.tcx, impl_substs);
-
         for item in self.impl_or_trait_item(impl_def_id) {
             if !self.has_applicable_self(&item) {
                 // No receiver declared. Not a candidate.
                 self.record_static_candidate(ImplSource(impl_def_id));
                 continue;
             }
 
+            let (impl_ty, impl_substs) = self.impl_ty_and_substs(impl_def_id);
+            let impl_ty = impl_ty.subst(self.tcx, impl_substs);
+
             // Determine the receiver type that the method itself expects.
             let xform_tys = self.xform_self_ty(&item, impl_ty, impl_substs);