CUDA Support

[rpep]

• Port energy kernels (easy)
• Port Demag (easy-ish)
• Port time integration (hard)

Unless we can get SUNDIALS GPU working, would have to ditch this and write own integrator. At the moment cufft is invokable only from host which stops this being possible even with new versions, as SUNDIALS wants integration function to run entirely on GPU.

[davidcortesortuno]

We also need to refactor all the energy classes into C to avoid calling Python to update the effective field. From the branch https://github.com/computationalmodelling/fidimag/tree/make-fidimag-great-again it should be easier to maybe have CUDA support in the future