Fix a performance regression in obligation deduplication.

Commit 8378487 from rust-lang#114611 changed the location of an obligation
deduplication step in `opt_normalize_projection_type`. This meant that
deduplication stopped happening on one path where it was still
necessary, causing a couple of drastic performance regressions.

This commit moves the deduplication back to the old location. The good
news is that rust-lang#114611 had four commits and 8378487 was of minimal
importance, so the perf benefits from that PR remain.

Fixes rust-lang#116780, rust-lang#116797.

compiler/rustc_trait_selection/src/traits/project.rs

@@ -1233,7 +1233,7 @@ fn opt_normalize_projection_type<'a, 'b, 'tcx>(
 
             let projected_term = selcx.infcx.resolve_vars_if_possible(projected_term);
 
-            let result = if projected_term.has_projections() {
+            let mut result = if projected_term.has_projections() {
                 let mut normalizer = AssocTypeNormalizer::new(
                     selcx,
                     param_env,
@@ -1243,14 +1243,14 @@ fn opt_normalize_projection_type<'a, 'b, 'tcx>(
                 );
                 let normalized_ty = normalizer.fold(projected_term);
 
-                let mut deduped = SsoHashSet::with_capacity(projected_obligations.len());
-                projected_obligations.retain(|obligation| deduped.insert(obligation.clone()));
-
                 Normalized { value: normalized_ty, obligations: projected_obligations }
             } else {
                 Normalized { value: projected_term, obligations: projected_obligations }
             };
 
+            let mut deduped = SsoHashSet::with_capacity(result.obligations.len());
+            result.obligations.retain(|obligation| deduped.insert(obligation.clone()));
+
             if use_cache {
                 infcx.inner.borrow_mut().projection_cache().insert_term(cache_key, result.clone());
             }