Fix some issues

[enra64]

This PR fixes some issues. It also, yet again, reformats everything - I did this so I could use a common formatter on my diverged fork from felixmusmann/jointly, making getting my fixes much easier. I've added the formatter (black) as a GitHub action, so hopefully in the future no more divergent formatting will be considered.

The list of issues this branch fixes is as following:

• by resampling the data for each column sequentially and creating a dataframe from that, we can avoid the large memory allocation required for resampling every column in a single dataframe. This change takes place in helpers/get_equidistant_signals.
• There is some issue with plotting the shake segments if the logger is set to INFO - probably something similar to the problem above, since a MemoryError gets thrown. However, since I didn't need those plots, I simply catch the error if it occurs and skip the drawing - this could maybe be made into an issue.
• I've made the start- and end-window seperately configurable, and added an error if the window lengths are longer than the data
• If start- and end-shake are detected to be the same, an error is thrown
• This branch should fix an off-by-two error that occured because the shift index after correlation was wrong. I think MSCL tried to fix that problem here, but that was more of a band-aid - after changing shift_in_samples in synchronizer to add 1 instead off subtracting it, the reference signal became much better synced, even without MSCL's fix
• If the threshold is smaller than 0 or greater than 1, an error is thrown
• If first or second shake is missing, an exception is thrown
• This PR adds a function to export pickled dataframes for each sensor indivvidually - this makes reading them afterwards much faster, and avoids the extremely sparse CSV that could be generated if the synchronization created many different data indices (up to 7*128=896 per second per sensor for me)