More optimizations

• Speedup backprop
• Fix gpu backend for different CUDA versions

juno_samples/matmul/src/matmul.sch

@@ -52,6 +52,8 @@ if feature("cuda") {
   infer-schedules(*);
   dce(*);
   //rewrite(*);
+  let out = outline(matmul@outer);
+  gpu(out);
   fixpoint {
     simplify-cfg(*);
     dce(*);