cuda: Fix wrong write combined flag

The buffer sent to CUDA needs write combined flag.  The buffer send
to the user does not need this.

This flag made Registration::apply() very slow in its memory read.

Author: xlz

src/cuda_depth_packet_processor.cu

@@ -535,7 +535,7 @@ private:
   bool allocate_cuda(Buffer *b, size_t size)
   {
     unsigned int flags = cudaHostAllocPortable;
-    if (!input)
+    if (input)
       flags |= cudaHostAllocWriteCombined;
     CHECK_CUDA(cudaHostAlloc(&b->data, size, flags));
     b->length = 0;