Fix continuous integration

[astrofrog]

Work in progress!

[astrofrog]

Looks like the difference is due to changes in the source positions:

position        y of </Sources/source_03567> y of </Sources/source_03567> difference          
------------------------------------------------------------
                -9.0529         -9.0529         1.77636e-15  

In some cases. We could also try and make the source names more reproducible to minimize diffs between HDF5 files but that's less important. So need to figure out why in e.g. test_ptsource_vertices_noclip the source positions are changing.

[astrofrog]

The numerical differences appear to be due to numpy/numpy#19478

[astrofrog]

Possible fix would be to use:

            export NPY_DISABLE_CPU_FEATURES=AVX512_SKX

but need to check how we would force Numpy to install from source.