Have the global-device-memory versions of make_unique and make_unique_region be the default

[eyalroz]

Right now, we don't accept cuda::memory::make_unique<T>, nor cuda::memory::make_unique_region - you have to use a subnamespace - one of device::, host:: or managed::. But - if you say:

cuda::memory::make_unique_region(my_device, 1234);

it's pretty obvious you want plain vanilla device-side memory. So, let's accept that.