Static frequency mask estimation from visibility data

[ljgray]

The overall goal of this PR is to implement a task to come up with a static RFI mask from visibility data directly, without having to rely on a hard-coded set of bad channels.

This produces a flagging metric by reducing over baselines using the existing ReduceChisq code as follows:

• The chi2 is computed over cross-pol baselines only, and excludes baselines < 5 meters.
• Weighting is done by dividing by baseline redundancy, which highlights faint frequency bands which are only otherwise visible in the high-delay chisq metric. I'll show some plots at some point.

Flagging is then done using a simple 1D MAD metric over the 1D median in time of the chisq metric. Only nighttime data is included in the median (the hook is implemented in a ch_pipeline PR). It uses an updated version of the arPLS baseline-fitting code, which is better able to handle the lack of hard-coded static mask.

This PR does the following:

• Updates containers.group_tasks to be able to handle multiple setup methods. This is still a bit fragile, but it works.
• Updated the scale_invariant_rank operation to be a bit faster (although I didn't use it in the end)
• Adds options to the MaskBaselines task to mask polarisation and to optionally combine the mask using an and operation
• Extends the ReduceBase task with a hook for custom weights
• Updates the penalized_least_squares implementation to support different weighting schemes. This implementation is a bit of a hack, abusing the __new__ class method to make it appear like a function, while still being able to subclass with custom hooks. I'm open to thoughts about how else to implement this.
• Replaces the old narrowband vis flagging task with this new static RFI mask task

I think the commits should be fairly easy to follow.

This code all works and is tested - I just haven't been able to thoroughly test how well the method actually performs on a wide selection of days