Investigate underestimated noise in coadds

Science users have reported that the noise model underestimates the actual noise in coadds by a factor of 2 to 3. My own investigations have shown that there is a low-level pattern in pfsArm data that emerges when coadding many exposures (even without resampling) that appears to be the origin of this. Investigate where this is coming from, and how we can remove it.