Add cadenced calibration

[Joshuaalbert]

We define two intervals: solution interval and validity interval. The calibration happens on an averaged down solution interval, and it is applied to validity interval.

How:

• When calibration happens it produces a set of gains specific to the data calibrated on.
• The shape of gains will thus match the averaged visibilities. [source, time=1, antenna, freq, 2, 2]
• In order to apply the gains to other visibilities, they must be expanded out again to match future visibilties.
• The natural structure to match against is time chunks, since averaging reduces time down. So we can repeat over time easily, or perform any other predict-type extrapolation.
• Thus we have logic that calibration happens only if validity interval has passed.

[Joshuaalbert]

Developed extended state space model to manage pushing the solutions. Not sure it'll fast enough. Will need to use JVP in several places to avoid materialising jacobians.