Core: cuda backend improvements

[GPMueller]

The speed on maxwell and pascal is roughly equal, which should not be the case. It seems that on pascal some quite trivial kernels take far longer than they should.
A cause for this might be that on pascal not all of a shared memory object is pre-fetched when it is accessed. This may cause significant slowdown in incremental accesses, i.e. in such cases arrays should be pre-fetched by calling cudaMemPrefetchAsync - but maybe cudaMemAdvise could suffice.

• cu_project_tangential
• cu_sum
• cu_set_c_a2

[GPMueller]

It seems the cudaMemPrefetchAsync has negative impact on the performance... Strangely it seems to significantly decrease the number of iterations after which the IPS drops.

Another idea is the following: each iteration, cudaMallocManaged is called 10x, which may indicate an unnecessary copy (missing reference? an = where it doesn't belong?).

[GPMueller]

It seems the removal of several cudaMallocManaged calls had significant impact: 23cd093

[GPMueller]

The removal of systems[0]->UpdateEnergy(); from Method_LLG::Hook_Post_Iteration() improves performance by another factor ~1.7 but it is unclear to me why.

[GPMueller]

The answer seems to be that the gradient and energy calculations for Exchange and DMI are all very costly due to the use of atomics.

A new scheme of the Hamiltonian is needed to make it better suitable for such parallelisations.

Related to #101 and #146

[GPMueller]

New schemes for the Hamiltonians have been implemented (8569ae5). #222 now tracks the implementation of the cuda versions of these.

[GPMueller]

Further performance improvements will, at my level of expertise, be algorithmic (see e.g. #311) and not in terms of better CUDA code. Therefore closing this issue.