Tiling optimizations

• Tile in cava to vectorize gamut fused channel loop
• Fix tiling in BFS to use correct loop order
• Tile further in BFS for vectorization
• Put init loop on GPU for BFS

hercules_opt/src/fork_transforms.rs

@@ -1741,6 +1741,7 @@ fn extend_fork(editor: &mut FunctionEditor, fork: NodeID, join: NodeID, multiple
                     control: new_fork,
                     dimension: idx,
                 });
+                edit.sub_edit(fork, tid);
                 let old_bound = edit.add_node(Node::DynamicConstant { id: *old_factor });
                 edit.add_node(Node::Binary {
                     op: BinaryOperator::LT,