Create a new "meas_extensions_ngmix" package, and add SFM and Forced plugins to approximate the PSF using a mixture of Gaussians using ngmix's E-M fitter.

There is example code to actually run ngmix in this mode in ngmix's README.md, but the tricky part is probably figuring out how to extract the outputs and write them to records. If at all possible, we should use the same format as modelfit_ShapeletPsfApprox, so either algorithm could be used to feed CModel (and other ngmix algorithms, when modelfit_ShapeletPsfApprox is configured to use only zeroth-order fitting).

This will also require figuring out the failure modes and defining the appropriate flags. Given that ngmix has no C++ interface, this work should be entirely Python, and hence it may be desirable to address DM-4009 somehow first.

This issue is almost certainly too large in terms of story points, but anything smaller doesn't really have a very well-defined end-point. Maybe subtasks?