Miscellaneous cleanup in CUDA code

Various changes to the CUDA code (and some common code as well) to make it simpler and less error-prone:

• Remove overly flexible operators of cu::HostMemory and cu::DeviceMemory
• Minor cleanup of auxiliary::Memory classes
• Let cu::*Memory classes derive directly from auxiliary::Memory
• Remove copy_htoh, copy_htod, and copy_dtoh from InstanceCUDA
• Use std::copy_n instead of memcpy
• Replace plain pointers to InstanceCUDA by std::unique_ptr