Fix eigen decomposition; new feature to store all optimization tries; and updates to documentation

[d-schindler]

I made a few changes to the code:

• Fix: Replace linalg.eig with linalg.eigh for the spectral decomposition used for computation of exponentials. We only apply spectral decompostion in the case of undirected graphs where linalg.eigh is the faster method (and we actually found an example where linalg.eig does produce incorrect results for an undirected graph).
• Feature: I introduced a parameter with_all_tries in the pgs.run() function that stores the information from all optimization tries in the results dictionary under an additional key "all_tries". For some applications it is interesting to look at the different optimizations produced for the same scale, not only at the best one that is selected.
• Fix: There are a few things I corrected/updated in the documentation.

[d-schindler]

I also renamed block_detection_curve in the optimal scale selection to block_nvi to make it consistent with our figure labels and the terminology in our software paper.

[d-schindler]

I added a figsize parameter to the plot_scan function, which is something I found quite useful.

[d-schindler]

@arnaudon , before we merge this PR we should probably update the online documentation of the module.

[d-schindler]

Further improvement of the documentation: I added the missing division by in the formula for MS with the continuous combinatorial Laplacian.

[d-schindler]

I've also added more references to the formulas for constructors. @arnaudon , I used reference [1] several times and wasn't sure if I should repeat it in every docstring or not.

[arnaudon]

I think I didn't put this here, so one can call figure before this function, and have more freedom