Can we add unimodal covariance to PPEs

[dehann]

Something like: ppe.mean ppe.covariance

all sorts of issues for multimodal, non-Gaussian cases, etc. The value "Covariance" is still perfectly valid, but then we would also need a way to support PPEs that give multimodal estimates (what are mean and covariances of top 3 distinct modes).

• This starts up all the GMM discussions, EM, Dirichlet Process, buffett/restaurant concentration parameter, etc.
  • PS, that's what we'd like to build out more of here: https://github.com/JuliaRobotics/KernelExpectationMaximization.jl