Use fewer delayed bugs.

For some cases where it's clear that an error has already occurred,
e.g.:
- there's a comment stating exactly that, or
- things like HIR lowering, where we are lowering an error kind

The commit also tweaks some comments around delayed bug sites.

compiler/rustc_const_eval/src/transform/validate.rs

@@ -117,18 +117,14 @@ struct CfgChecker<'a, 'tcx> {
 impl<'a, 'tcx> CfgChecker<'a, 'tcx> {
     #[track_caller]
     fn fail(&self, location: Location, msg: impl AsRef<str>) {
-        let span = self.body.source_info(location).span;
-        // We use `span_delayed_bug` as we might see broken MIR when other errors have already
-        // occurred.
-        self.tcx.dcx().span_delayed_bug(
-            span,
-            format!(
-                "broken MIR in {:?} ({}) at {:?}:\n{}",
-                self.body.source.instance,
-                self.when,
-                location,
-                msg.as_ref()
-            ),
+        // We might see broken MIR when other errors have already occurred.
+        assert!(
+            self.tcx.dcx().has_errors().is_some(),
+            "broken MIR in {:?} ({}) at {:?}:\n{}",
+            self.body.source.instance,
+            self.when,
+            location,
+            msg.as_ref(),
         );
     }