Astroalign

Bug fix:

* Images with small number of sources < 5 would not work with `find_transform` and `register`. (see 36)

This only tests the case when `source` and `target` are input as a list of (x, y) positions (as opposed to the image array).

Separate cases for 3, 4, 5, and 6 sources.

Note: Even though this version may still work with Python 2.7, this version drops support for testing Python < 3.

What's new

* `register` now accepts as input images Astropy's [`NDData`](https://docs.astropy.org/en/stable/api/astropy.nddata.NDData.html) objects (this includes the CCDProc's [`CCDData`](https://ccdproc.readthedocs.io/en/latest/ccdproc/ccddata.html) subclass) as well as Numpy's [`ndarray`](https://docs.scipy.org/doc/numpy/reference/generated/numpy.ndarray.html).

* `register` returns a `footprint` boolean image, `True` for masked pixels with no information.

registered_image, footprint = aa.register(source, target)


* New `fill_value` parameter: convenience argument to fill out untouched areas on the registered image.

registered_image, footprint = aa.register(source, target, fill_value=-99999.99)


* For data objects like `NDData` or Numpy masked arrays, `register` now propagates the mask if any.

**Note**:
This version drops support for the deprecated functions `align_image` and `find_affine_transform`.

More information and examples in the [documentation](https://astroalign.readthedocs.io/).

Version 2.0 will return the footprint of the transformation, and deal with data objects with masks (CCDData, Numpy masked arrays, NDData) in a more consistent way.

Minor Fixes

Release Notes

* Improved tests

Release notes

* Removed unused `astropy` dependency.
* Fixed issue 20 in which skimage raises error if image is of type f4 ('float32').
* Added docs for `estimate_transform` and `matrix_transform`.
* Added more tests.
* Bug fixes.