More optimizations

• Speedup backprop
• Fix gpu backend for different CUDA versions

juno_samples/rodinia/bfs/src/gpu.sch

@@ -47,5 +47,5 @@ let collect_body = outline(inner);
 fork-tile[1024, 0, false, true](traverse, collect);
 fork-split(traverse, collect);
 
-unforkify(init);
+unforkify(init_body);
 gcm(*);