GPU Tracer: Add peer-to-peer memcpy annotations (#11795)

In order to trace memory copies between devices with more recent versions of
CUDA, the GPU tracer must capture peer-to-peer (i.e. device-to-device) memory
transfers (e.g. cuMemcpyDtoD, cudaMemcpy(., ., ., cudaMemcpyDeviceToDevice)).
Add handling to capture these.