"Loose" (Ungridded) Visibility DataConnector

[iancze]

Is your feature request related to a problem or opportunity? Please describe.
The gridding process within the current Gridder object makes assumptions about the data that, while fast to compute, obscure some potential calibration and processing subtleties.

1. makes the assumption that the visibility function is constant across a u,v cell. This is technically something that could be improved within the GriddedDataset framework.
2. by collapsing all of the loose visibilities within a u,v cell to a single, averaged summary statistic (the gridded mean and gridded weight) on that cell, we lose the ability to visualize and assess differences between the loose visibilities. For example, we might want to examine if the loose visibilities (or residuals to them) have any trends with time of observation (suggesting a calibration error).

Describe the solution you'd like
Provide a functionality to interpolate the dense Fourier grid to the loose visibilities, calculating a model visibility for every "loose" datapoint. This will allow us to implement things like spectral covariances (#18) and data self-calibration (#24).

Rather than bringing back the tried and true prolate spheroidal wavefunctions in earlier versions of MPoL and DiskJockey, I think it makes sense to try out torchkbnufft which uses Kaiser Bessel functions (similar to FINUFFT). MRI and medical imaging in general has driven the development of these tools, and they're likely to be better than something that we would code from scratch.

Describe alternatives you've considered
It's possible that PyTorch conv2D layers could be adapted for the convolutional resampling, either from dense grid -> loose visibilities or even loose visibilities -> dense grid.