Noise sampling in curve simulation

[jrgant]

Originally misapplied the Brooks et al. noise sampling scheme.

Problem
For a given curve simulation, I use the noise from the trendfilter prediction that corresponds to the randomly selected template shape (i.e., from the same empirical season).

Fix
Brooks et al. actually choose both a shape and a noise value at random for each curve simulation.

Don't think this will matter much, but should fix to more closely approximate their procedure.