Improve encoding memory bound estimate

[jeromekelleher]

Currently we just return the maximum size of a variant chunk, but there's also some overheads. Some workable rule of thumb would be good here that works across dataset sizes.