AutoMetaCal performance

[andrevitorelli]

Currently, amc handles 100k galaxies in 10min in a single Titan Xp.*
This is about the same ngmix takes with a 2x Xeon Silver 4214 (using 48 threads) setup w/ 125GiB RAM.*

However, the Addons>ResamplerGrad conversion to a tf graph returns the warning:
WARNING:tensorflow:Using a while_loop for converting Addons>ResamplerGrad

...which I think might be affecting performance greatly. This issue is to ask @Dr-Zero to look into this.

This is also where I will report performance comparisons in general.

*In this test we use simple fixed-width gaussmoms for both codes. I want to do a model fitting run with @b-remy and see how that compares.