Support for mapped/unified memory

[psychocoderHPC]

It is useful to have in some cases support for mapped/unified memory.
This allows the user to allocated host memory and use this memory directly on the GPU without waisting useful GPU memory.

In CUDA there are two kinds of host memory which can be addressed from the GPU without the need of an explicit memcopy.

• portable memory https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#portable-memory
• managed memory https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#mapped-memory

A open question is if such a feature is also supported by other platforms like AMD gpus or maybe later FPGs.

[BenjaminW3]

Unified memory already in #296

[BenjaminW3]

Mapping CPU memory into GPU memory space should aready be supported but is untested:
alpaka::mem::buf::map(buffer, targetDevice);
The device pointer can than be accessed via auto ptr = alpaka::mem::view::getPtrDev(buffer, targetDevice).
I do not know if this works.

[psychocoderHPC]

@BenjaminW3 thx I will have a look into alpaka unified memory.