Optimize the speed of calibration in time-domain
When running the pipelines in time domain, multiple xarray datasets are constructed and merged each time, which is a costly operation. In calibration of CDM for example this brings almost 50% longer calibration. Idea to be explored: Maybe it is better that the the basic output of the time-domain pipeline is numpy arrays in a dictionary. This should take less time to construct. Then xarray datasets are created when needed, as late as possible. Xarray dataset still remains the main output for every running mode.