Fine-tune scheduling of parallel reductions

• Fine-tune GCM to allow nested fork-joins involving parallel reductions.

hercules_cg/src/gpu.rs

@@ -1521,6 +1521,8 @@ extern \"C\" {} {}(",
                 let collect_variable = self.get_value(*collect, false, false);
                 write!(w, "{}{} = {};\n", tabs, define_variable, collect_variable)?;
             }
+            // Undef nodes never need to be assigned to.
+            Node::Undef { ty: _ } => {}
             _ => {
                 panic!(
                     "Unsupported data node type: {:?}",