mapped memory allocation currently insists on CU_MEMHOSTALLOC_DEVICEMAP

[eyalroz]

Currently, our cuda::memory::mapped::allocate() function always allocates with the CU_MEMHOSTALLOC_DEVICEMAP flag - meaning it expects to get the same address for the device and the host. But - this whole namespace is for when we have a pair of addresses...

Either we support real pairs, in which case we support older GPUs as well (which is the earliest CPU to provide the same address? Kepler? Maxwell? Can't be later than Pascal) - or we rearrange this namespace somehow.