Differences between mesh generation on Win and linux result in fragile CI

[krcools]

In test/test_hh3d_nearfield, the spherical mesh generated by gmsh is different on WIndows and linux for the specific parameters passed to CompScienceMeshes.meshsphere. This can result in error bounds set in a Linux environment to be too tight for Windows.

It would be better to load mesh files from an asset subdirectory of test than to generate the mesh on every run. Not only is this more robust, it also results in fewer non-test related CPU time.

[sbadrian]

We got a student working on this topic.

I think the default behavior should change to the Julia code, but maybe we would like to retain the possibility to generate a sphere or cube mesh via GMSH? We could add a keyword argument to accommodate for this.