Multi-GPU parallel beam code

The rationale to not have this before was that the parallel beam case was trivially parallelizable, i.e. you could call 4 instance of TIGRE, instead of spreading a single instance over 4 GPUs, and many people preferred that on the syncrotron tomography side.

With the introduction of #242, this is not require anymore. One can simply give a single GPUid to each instance of TIGRE, or can instead run all the code in many GPUs from a single thread.

Therefore, the parallel beam source should be made multi-GPU compatible.