- Use direct prediction for the measurement equation, rather than FFT and
degridding. This could potentially be slightly more accurate, although the
images are not noticeably different. The old behaviour can be enabled with
``--degrid``, which could be faster if there are very large numbers of CLEAN
components.
- Fix some bugs that could cause CUDA errors.
- Fix a bug that would cause a crash for some image sizes.
- A very small correction to parallactic angle calculations.
- More efficient loading of katdal datasets.