Unify compute_isi_violation docs and add UltraMegaSort2000 citation

[chrishalcrow]

An attempt to close #2050

Two things:

1. Unify the notation between isi_violation and rp_violation in the docs so that the functions are easier to compare.
2. Get a correct citation for isi_violation: the current implementation is not actually from the Hill paper, it's a simpler function originally from UltraMegaSort2000 (https://github.com/danamics/UMS2K, https://github.com/danamics/UMS2K/blob/master/quality_measures/rpv_contamination.m) which is also written by Hill. I think citing this directly is much less confusing (see more confusion here: Contamination ratio #1973, and see https://github.com/cortex-lab/sortingQuality/blob/70c8659adc60484434be828d617e16eb83e94cca/core/ISIViolations.m#L17 for another discussion. Here our isi_violation formula is "the approximation previously used" which "wasn't accurate to the Hill paper" )

[JoeZiminski]

Looks good to me! Some small suggestions

[JoeZiminski]

great stuff!

[zm711]

Nothing to stop merging, but a couple potential tweaks around the edges from my perspective.

[zm711]

This isn't 100% clear. Worse being what exactly... "As it grows past 0.xx it begins to diverge from xx" something like that based on llobet's comments in the issue.

[chrishalcrow]

Mmm. I was deliberately vague because the success of the approximation depends on lots of factors: when the contamination is large is it because there are LOTS spikes near the same time a few times, or many violations of where one spike violates only one other spike? The Llobet comment is on one specific example (uniformly generated random spikes) so you shouldn't take precise numbers from it.

All the models agree when $n_vT/N^2t_r$ is small (much less than one), so we could say that the interpretation only works in that situation.

[zm711]

That's fair. Okay so maybe something more like:

This intrepretation is good when the ratio is small, but as the ratio approaches 1 it is a less reliable indicator (or something like that? ). I know the ratio can go above one so maybe we don't even want that. But "becomes worse" is he vague part for me. ie becomes worse at what?

[chrishalcrow]

I've realised that's also not true: you can cook up examples (where each spike is contaminated by exactly one other spike) where the ratio is a reliable indicator. So we can only really say: sometimes this is a good indicator, sometimes it isn't (like most of the quality metrics!). I suggest we refer to the Llobet paper, which discusses some of the different models? I've had a go.