Experimental Python port of Kilosort 2

[rossant]

This pull request provides a merge of the rossant/pykilosort repository, containing an experimental Python port of Kilosort 2, into @marius10p's master Kilosort2 repository. The merge was done with git following these instructions.

The Python version is found in the pykilosort/ subdirectory. This version is not ready for production yet. The README of pykilosort hasn't been merged at the moment.

• Line-by-line port from MATLAB into Python of the core implementation
• CUDA kernels have been kept completely untouched
• CuPy has been heavily used to match MATLAB's heavy use of GPU array and matrix operations
• CuPy was missing some features used by the MATLAB implementation (eg the median), most importantly an efficient GPU implementation of lfilter(). Effort has been done, in collaboration with @oliche, to make a scalable, memory-efficient, and fast FFT-based convolution of arbitrarily long multichannel time series.
• Development has been done using 2 excerpts datasets of 100s and 1000s from a Neuropixels 384-channel recording
• On these test datasets, performance was grossly but not perfectly similar in terms of sorting quality and time performance
• Validation with a sort of ground-truth dataset has not yet been performed.