Numpy-groupies

Performance improvements

What's Changed
* Fix versioneer configuration. by sebastic in https://github.com/ml31415/numpy-groupies/pull/71
* Fix pytest failure without numba. by sebastic in https://github.com/ml31415/numpy-groupies/pull/72
* Move to micromamba by dcherian in https://github.com/ml31415/numpy-groupies/pull/76
* Use faster group_idx creation when axis == -1 by dcherian in https://github.com/ml31415/numpy-groupies/pull/77
* Some optimizations for numpy aggregations by dcherian in https://github.com/ml31415/numpy-groupies/pull/78

New Contributors
* sebastic made their first contribution in https://github.com/ml31415/numpy-groupies/pull/71

**Full Changelog**: https://github.com/ml31415/numpy-groupies/compare/v0.9.20...v0.9.21

Fixes some issues for argmin and argmax

- Fixes an issue for non-default fill_values for mean/std/var.

This version comes with a couple of new functions for the numba implementation of aggregate:
- `argmin`
- `argmax`
- `cumsum`
- `cumprod`
- `cummin`
- `cummax`

The numba implementation now also handles jitting of arbitrary reduce functions, though the speed is not that impressive as it might sound.

`sort` was modified and sped up, so that it produces a groupwise sorted array in the same size of the input, and not an array of arrays anymore. This should be the more useful mode of operation.

This version is finally fully tested and compatible with python3. Note that the weave implementation is not available in python3, as python3 is unfortunately not supported by them. We recommend the numba implementation for new projects!