Apply row concept to time and antenna axis

[sjperkins]

Note that this applies to the new version in the dask branch.

In the dask version a unique time, as well as a row dimension exist. These two dimensions are related to each other in that multiple rows (baselines) are associated with a unique timestep. Then, multiple rows associated with multiple timesteps can be processed by a single CUDA kernel.

This is useful for cases where baseline averaging is performed. While correlators measure a visibility along all baselines at the same time, baseline averaging produces different time centroids. By computing row visibilities for multiple timesteps, instead of baseline visibilities for multiple timesteps, compute is saved.

This can also be extended to the antenna dimension. As the number of baselines per timestep is reduced in baseline averaged data, so too is the number of antenna. By only computing antenna values for only those present in baseline for a timestep, we can avoid computing antenna values for all antenna.

[sjperkins]

/cc @o-smirnov @bennahugo