make peeling more reliable

[mweastwood]

Peeling is currently very sensitive to unmodeled flux. Especially RFI sitting on the horizon.

[mweastwood]

There appears to be several components to this:

• getting a good starting guess (ie. a beam model build in a beam model #23 + better source models)
• modeling the unmodeled flux (especially the galactic plane)
• RFI suppression
• amplitude-only calibration

The amplitude-only calibration will help guarantee that you are removing flux from the expected direction on the sky. If each source is allowed a full complex gain calibration, then flux can be removed from almost anywhere. (Edit: this is not useful in practice, what you really want is a prior on the phase of the solution so that it's not too far from zero)

[mweastwood]

• Warn if the peeling solutions have large phases (ie. flux is being subtracted from elsewhere on the sky)
• Sort sources by expected flux before peeling (in the meantime, sources are peeled in order from the sources.json file)
• Output peeling solutions (add an option to save peeling solutions as numpy arrays #44)

[mweastwood]

New idea:

• Add a constraint that minimizes variance in the peeling solutions.

In order to move a source, you need to apply a phase gradient across the array, so dialing up this constraint should put an end to that.