Now the CUDA compute capability is not hard-coded anymore.
Setting

--with-gpu-compute-capability=value

set the nvcc compiler option "-arch value".

A simple sanity checking is done to ensure that "value" is not
completely unreasonable