Improve performance and scalability of calibration in ICAL

Copied from Miro (see: https://miro.com/app/board/uXjVKZ900uo=/?moveToWidget=3458764589180836457&cot=14 ): We want our ICAL pipelines to be able to keep up with the Low and Mid telescopes at AA2 scale, i.e., these pipelines should take at most twice the duration of the processed observation. We are currently a significant factor away from that goal.

• The most recent benchmarking results for the Low pipeline show that the workload for calibration is increasing each cycle due to the increasing complexity of calibration due to an improving sky model. This is confirmed by the timings from the Rapthor pipeline which also include a 4th and 5th cycle with the latter using 100% of the data instead of ~20% of the data. This means that calibration could well become the main bottleneck of the pipeline.
• The most recent benchmarking results for the Mid pipeline show that the predict stage in WSClean does not scale to multiple node (it actually becomes slower). This needs to be investigated and addressed.

1. Introduce image-based prediction in calibration by using the w-gridder (DUCC) in a DP3 Predict Step.
2. The memory usage of DP3 is considered carefully, so that DP3 does not start swapping. A mechanism is thus introduced to balance gridding efficiency (favours large time and frequency intervals) and memory usage (favours short time and frequency intervals).
3. The bad scaling of the predict stage in WSClean in the Mid pipeline is addressed.