implement copy_to_peer

[smedegaard]

The problem

hipMemcpy docs states the below about multi-gpu and peer-to-peer copy.

types::memory::copy_to_peer()

It supports memory from

• host to device
• device to host
• device to device
• host to host

The src and dst must not overlap.

For hipMemcpy:
the copy is always performed by the current device (set by hipSetDevice).

For multi-gpu or peer-to-peer configurations:

• it is recommended to set the current device to the device where the src data is physically located.
• For optimal peer-to-peer copies, the copy device must be able to access the src and dst pointers (by calling
• hipDeviceEnablePeerAccess with copy agent as the current device and src/dest as the peerDevice argument.
  if this is not done, the hipMemcpy will still work, but will perform the copy using a staging buffer on the host. Calling hipMemcpy with dst and src pointers that do not match the hipMemcpyKind results in undefined behavior.