Clean implementation of dipole-dipole interactions using FFT convolutions

[MSallermann]

An Implementation of a convolution algorithm to speed up the calculation of the dipole-dipole interactions. On the CPU it ships with kissFFT but provides the options to use FFTW in core/CMakeLists.txt. On the GPU it uses cuFFT. The interface to the different libraries is wrapped in core/include/engine/FFT.hpp so that the calls in the rest of the code are (mostly) library-independant.

TODO:

• [x] Get performance back by removing idx_from_tupel calls in .../Hamiltonian_Heisenberg.cpp
• [x] Speed up the cuda CU_ptwise_Mult kernel as it slows down the entire code
• [x] check all prefactors for mu_B (see #444 )
• [x] implement an option to save the dipole-matrices instead of throwing them away after they are fourier transformed (useful for hessian)
• [x] re-implement the approximation of periodic boundary conditions for CPU
• [x] implement the approximation of periodic boundary conditions for GPU
• [x] maybe add an option to configure how many periodic images should be taken
• [ ] re-implement the use of Eigen to perform pointwise multiplication (?)
• [x] make it possible to turn DDI on/off
• [x] make it possible to switch DDI method (e.g. cutoff, FFT, etc)
• [x] make the CPU code skip redundant entries in the pointwise multiplication for both kissFFT and FFTW
• [x] write a unit-test (?)
• [x] more OpenMP use
• [x] adapt improved memory layout to CUDA
• [x] fix a silly mistake where there are actually ~n_cell_atoms^2 reverse FFTs performed intead of ~n_cell_atoms
• [x] naming conventions overhaul
• [ ] reduce the number of kernel launches if there are multiple atoms per unit cell
• [x] also write the inputs for the fourier transform of the dipole matrices in a kernel to reduce start-up time
• [x] correct handling of single/double precision for fftw/cuFFT (currently fftw will always use 'double' and cuFFT will always use 'float')
• [x] automatically link fftw_openmp / improve find_fftw.cmake
• [ ] make the fourier transforms inplace to save memory

[coveralls]

Coverage decreased (-0.6%) to 84.412% when pulling 212dc3d526edabaefedab7826560255b58a6d8cd on MSallermann:feature-DDI-FFT-clean into e82976901e22c18a00fee80fdac607790ac26878 on spirit-code:develop.

[codecov[bot]]

Codecov Report

Merging #446 into develop will increase coverage by 1.77%. The diff coverage is 89.68%.

@@            Coverage Diff             @@
##           develop    #446      +/-   ##
==========================================
+ Coverage    49.42%   51.2%   +1.77%     
==========================================
  Files           88      90       +2     
  Lines         9452    9773     +321     
==========================================
+ Hits          4672    5004     +332     
+ Misses        4780    4769      -11