The loops made it hard to find where these keys were created, and there
are helpers to create flux and centroid keys with their associated
uncertainties automatically.
Using NO_UNCERTAINTY for centroids, to make it explicit that we don't
have uncertainties for these values right now.

* Naive* are purely examples and shouldn't be used in production
* Peak* are unnecessary with Psf and Dipole models and mostly for
testing purposes
* PsfDipoleFlux is slower than PsfDipoleFit (this task).
* PsfDipoleFit provides a dipole classification.

PsfDipoleFlux is the older, slower algorithm, but it makes a useful
check on the new PsfDipoleFit.

* Enforce plugin order, so that DipoleFit can fall back on SdssCentroid
for non-dipoles.
* DipoleFitPlugin can run at "centroid" order, because it simultaneously fits
centroids and fluxes.
* Switch centroid slot after SdssCentroid, so that the "best" centroid
comes from DipoleFit (even if it's just copied over).
* Switch dipole centroid field names to better match centroid slot
convention (foo_x/foo_y, for plugin foo).
* Rename DipoleFitTask default name to match Task name convention.

Cleanup tests to pass with the better centroids:
* Loosen flux tolerance.
* Remove test of "unphysical" sources that relied on bad centroiding
pushing the sky sources off the image.
* Remove tests that relied on old centroider behavior when measuring
on unmerged footprints.

Refactor non-dipole "error" handling:
* No need to flag sources that were not fit as failures, they're already
marked as not dipoles.
* Handle "too large" footprints as failures separately from other non-
dipole sources.
* Remove redundant flag setting: some flags either default to False, or were
set just above.

We want the debug logs to appear if the loglevel is DEBUG.
Plot display was broken because `show()` never got called!