Fast EM2D incorporated into ElM2D

[sgbaird]

This is a faster implementation of the generation of ratio and mod-petti vectors, as well as the computation of the pairwise distance matrix. I was hoping to get dist_matrix.py and related routines incorporated into one of the Numba repositories (see numba/numba#7377), but that probably won't happen for a while for various reasons. There is a non-CUDA and a CUDA version, with the non-CUDA version as a default.

The results of test_ElM2D.py:

Fitting mod_petti kernel matrix
Constructing distances
[fit-wasserstein]
Elapsed: 4.08646

Fitting mod_petti kernel matrix
Constructing distances
Parsing Formula
100%|██████████| 5000/5000 [05:31<00:00, 15.10it/s]
Constructing joint compositional pairings
100%|██████████| 4999/4999 [00:00<00:00, 5500.21it/s]
Creating Process Pool
Scattering compositions between processes                     and computing distances
  0%|          | 0/4999 [00:00<?, ?it/s]
Distances computed closing processes
Flattening sublists
[fit-network_simplex]
Elapsed: 718.77971