Sc3s

Main consensus clustering now allows passing of multiple `num_clust` parameters. The subfunction `strm_spectral` now returns the coordinates of microclusters, instead of consolidating into macroclusters directly. This allows the microclusters to be reused for the `num_clust` values, saving compute time.

Also fixed some random and miscellaneous bugs. A notable one is getting the streaming to work properly with the SciPy sparse matrix format.

Parallel executions of different k-means initialisations for each dimensionality reduction parameter `lowrankdim` now utilise the same Laplacian approximation, and hence the data is only streamed once for each parameter.

Also rewrote the code for the binary matrix construction, such that it can work with individual clustering runs/modularities more robustly, regardless of the number of unique clusters each of them has.

First official preproduction version, with minimal documentation for core function added.

**First working prototype, where all main functions modularised**

It is now possible to run the consensus clustering a single command from a `scanpy` script.
py
import sc3s
sc3s.tl.consensus_clustering(adata, num_clust, lowrankrange, n_parallel)


However, the parallel executions of different k-means initialisations for each dimensionality reduction parameter do not yet utilise the same Laplacian approximation. Currently, these are different streaming orders where the stream is restarted.

This will be changed in future releases, so the data is only read once (for each `svd`), meaning that the ordering of cells into the stream will be the same across all the k-means initialisations that are tested. We still need a new Laplacian for each parameter of dimensionality reduction.

Update tag to 0.0.2 to try resolve renaming issues.