A complex bug with y_scales and multi-channel data

[sg-s]

Problem: different channels have different scales, and combining them together (or even normalziging one channel across time) needs some scale to work with.

This is what is stored in the common.y_scales array. However, this is computed once, at initial file load, which means it may be wildly off if the data in the first file is filled with nonsense.

Workaround: update it when using mulit-channel dimensionality reduction