Add svd_solver to phase 1 sorter parameters

[chrishalcrow]

In #50, we discussed ways to speed up the compute_pca_features function in phase 1, which was the bottleneck for applying mountainsort5 my data.

When experimenting, I found that using the covariance_eigh algorithm for decomposition (https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html) sped things up a lot! This PR allows users to specify the phase_1_svd_solver parameter, but keeps the current default.

I don't have any solid benchmarks for saying that covariance_eigh works fine - but it works fine for my data!

Note that this does not get passed to isosplit6_subdivision_method because it takes a smaller subset of data, and the default method runs quickly.

If something like this gets merged, I can update the spikeinterface side too :)

[magland]

This looks great @chrishalcrow !

Assuming we believe that folks should use this going forward, without wanting to affect existing pipelines, I'm wondering if you would mind updating the docs to include covariance_eigh in all the examples? You could even make a comment about the speed improvement.