Implementing KNN referencing and a backup referencing mechanism for common_reference

[zzhmark]

PR based on this issue: #4193

The new features are designed for ungrouped local referencing.

Changes in parameters:

• reference: include knn as an option
• backup_reference: allowing global, single and knn (local is unstable) to take effect when local referencing cannot get enough #channels, default as global
• nneighbors: used for knn referencing, i.e. #channels nearby used for referencing
• backup_thr: when #channel found locally is below this, turn to backup_reference, default as 1, i.e. no channel found. (we may also consider None to force explicit assignment)

Changes in workflow:

• Verification in parameters. Since reference and backup_reference both need preprocessing, all the elif are turned as if.
• Removing assertion of no local reference found as it would be judged dynamically.
• Add knn preprocessing in init (similar to local).
• Changing the ungrouped part in get_traces.
  • Defining each reference method as local functions to call.
  • In local, setup if else to choose backup when #channels not meeting the requirement.

[zzhmark]

Here is an example ephys data I used for testing to see the differences. KNN is working better than global. The backup mechanism is verified by setting the local radius as (50, 51) (no channel would be found) and its backup method is global, so backup car is getting the same result as the global car, and it's different from the proper local car (50,200).

[alejoe91]

Hi @zzhmark

If I'm reading the implementation correctly, you are not discarding channels within the inner radius when computing nearest neighbors. I think this is the wrong behavior: if you rereference to closed-by channels, spiking signals will be essentially killed.
I think that we should still use the local radii, and keep the knn as backend when no channels are found. I think that was the original idea from #4193: to fix the case when local finds no channels. How does that sound?