First release running on Keras 3, and the first that works on a current
scientific Python stack. Requires Python 3.10 or newer and TensorFlow 2.16 or
newer.

The shipped DELIGHT_v1.h5 was written by Keras 2.7 and stored the network's
rotations, flips and concatenations as TFOpLambda layers, which Keras 3 cannot
deserialize, so the library had stopped working anywhere current. The
architecture now lives in delight/delight/model.py and the trained weights ship
as DELIGHT_v1.weights.h5.

The weights are unchanged and so are the predictions. Against a reference run of
the pre-migration code under TensorFlow 2.15 / Keras 2 / NumPy 1.26 / pandas 2.1,
the multi-resolution images, source extraction and sextractor coordinates are
bitwise identical, the network output agrees to 7.6e-06 px and the predicted host
coordinates to 7.3e-11 deg.

Also fixed: pkg_resources, matplotlib.cm.get_cmap and ndarray.newbyteorder, all
removed from their upstreams; a missing channel axis that Keras 2 used to add
implicitly; download() returning before its last batch of threads finished; and
plot_data() passing a masked array to np.percentile, which broke every
doplot=True path and had been broken before this release too.

Packaging moved from setup.py to pyproject.toml. pandas and scipy were imported
but never declared. The project had no tests; it now has 35.

Published artifacts:
  astro_delight-0.1.0-py3-none-any.whl
    sha256 b7f6a8884a9721151f997e34d7517ec6412e06aedec8ad9689f5c1edd884ad2f
  astro_delight-0.1.0.tar.gz
    sha256 3a0646424eb341e0dc6cc0974f140e9ae11a4a837336a29659dcd13a937dd0f0