Add percentile normalization

[timonmerk]

An option could be added in the nm_settings.json s.t. raw_normalization_settings / feature_normalization_settings have an argument use_percentile, lower_percentile, upper_percentile which limits the mean/median/z-score data distribution and addresses the effect of outliers.

See discussion in #285

[toni-neurosc]

Hi Timon, I was taking a look at those RuntimeErrors that were appearing in the normalization and I remembered this issue. Is this still interesting? It's easy to implement, but I'm not sure if do it as you say in "normalization_settings" or have it as an extra preprocessing step that you can add to the "preprocessing" list.

Also, in the discussion #285 you said that Commonly we face the problem that there are movement/cable noise artifacts that exceed many standard deviations of the data. If the artifacts deviate in many sigmas, then shouldn't the "clip" parameter of the normalization settings take care of that?

As long as you have z-score as the normalization method, the clipping is done in terms of standard deviations, the problem is what happens you're using any of the other normalization methods, or you're using no normalization at all, which seems to be the case you described in your message. In this case, it would be kinda useless to put the use_percentile option in the normalization settings, if you still skip normalization.

For this reason, I propose adding it as an extra normalization step, like remove_outliers or clip_percentiles, whatever you think is more descriptive. It's either that, or add a new dummy normalization method none that does not modify the data, so that you can use normalization with a use_percentile parameter without actually normalizing, but i feel it's too convoluted.