Perform tests to determine best default metacal reconv method

[esheldon]

gauss has been the default because it doesn't require any rng as input, so easier for newbies to use. However now that we know we need an rng in any case to fix the noise, this is not a benefit.

We use fitgauss in production because I imagined it would be more robust to elliptical PSFs but this was always just a hunch.

Develop some tests to see when these methods break, and which can provide more precise results.