Astroalign

Patch release with repo improvements

Change log

- Adds tea protocol
- Minor changes to `README`
- Fixed bad paths on `MANIFEST.in`
- Removes `setup.py`

In this release:

* Fixes compatibility with Numpy >= 2.0, by removing sep and replacing it with sep-pjw
* Drops support for python 3.8 that reaches EOL tomorrow.
* Adds 3.12 and 3.13 to CI testing

Minor release version to retroactively restrict Numpy version < 2.0 to continue using SEP.

This is released just in case there is a need to keep legacy software.

**Full Changelog**: https://github.com/quatrope/astroalign/compare/v2.5.1...v2.5.2

In this release:

- Mask in source or target image passes over the mask to source detection. Now is possible to exclude regions outside the region of interest ROI to find the match between images.
- Use of numpy's new random number generator API. This requires numpy version to be at least 1.17.
- Updates to README.md.

This release:

- Gets rid of `ez_setup.py` and uses `pyproject.toml` instead.
- Lazy loads `estimate_transform` and `matrix_transform` from scikit-image for faster import time.
- Drops support for python 3.6 and 3.7 from unit testing.

In this release:

Improved RANSAC algorithm

* The new `ransac` will exhaust all possibilities in a pseudo-random way. The exploration is done on a **single** (fixed) realization of all possible matches and it will exhaust the list of candidate matches before failing.
* `ransac` now iterates 3 times searching for the maximum number of matches before returning. This ensures (almost certainly in a statistical sense) that the results will be the same between calls.
* The new `ransac` is virtually deterministic, despite its stochastic search nature.