create approximate multinomial models

[rmcminds]

in principle, observations associated with larger counts should have more precise estimates for their relative abundance. at some point, this precision must become good enough that we truly can forget about the multinomial likelihood, fix the log relative abundance estimate, and treat it like a multivariate normal problem. But at what threshold does that make sense?

given a chosen threshold, implementation should be easy. instead of using a full matrix of 'latent_abundance' parameters, estimate only those that belong to observations with low counts. using multinomial likelihood on full vector of fixed and estimated parameters should still work, but may be more efficient to pool all fixed params into a single unit?