Python bindings in mixed CPU-gpu environment

[Yurlungur]

We may sometimes wish to build the python bindings on a system where CPU and GPU both are available. In this context, the python calls must still be host-only (because pybind/numpy are), but the device calls should be device-capable. This requires some tweaks in the python module.

[Yurlungur]

Note that on APU systems and similar systems with shared memory, this isn't an issue. But it is an issue on systems with a split memory space.