Improve MAPElites performance using torch_scatter

[JakeF-Bitweave]

Would you be interested in contributions to re-work MAPElites to use torch_scatter rather than the vmaped extended_population x feature_grid operation?

The general gist is:

• Quantize features and round them to longs
• Get the indexes for each elite for each volume using scatter_max (which returns argmaxes)
• Map the evals and values from the extended population to the population using the indexes