cosmic ray finder is faster uses less memory with numba

[julienguy]

Solves issue #612 . Reduced memory usage by a factor 4 using numba in the cosmic ray finder.

[sbailey]

Tested on an r5 image on cori, I find 20.5 sec -> 12.5 sec, and 2.2G -> 1.2G. The pixels, ivar, and readnoise are identical, while a tiny fraction of the masked pixels changed their mask value. Most but not all of those were edge pixels; there were 9 interior pixels whose mask value changed. That's good enough for me, but holler if you want to check those for completeness.

[sbailey]

For the record: Julien pointed out that the flat lamp exposures have higher S/N and the original algorithm used more memory in that case, so his 4x savings in memory isn't an exaggeration: it is indeed a 4x win for flats, which is huge because that will enable us to run all 30 cameras simultaneously.