Add the ability to fit baseline base instrumental noise terms

[rohandahale]

Currently, Comrade uses a fixed way of defining Likelihood based on the data products. Moreover, all the model parameters are inside the model visibilities while the sigma is fixed based on the errors of the data products. If likelihood function was made modular, we can do variability modeling and multi-frequency modeling.

[ptiede]

Let's talk about this carefully. Right now, the standard variability noise model is not something I am particularly fond of. Now maybe we do want to include some ability to fit for residual systematic errors, but I'd like to be a bit careful there. But if you want to take a shot at this in the new Comrade 0.10, I am very open to it.