Try out the specialized distributions that can be dispatched to be mean zero (i.e., passing in just the diagonal covariance matrix, or uses the FillArrays.Ones, etc.)
See if we can avoid passing back a DistributionsAD type and go back to a Distributions one.
Note that MvNormal(Diagonal([0.1, 0.2])) |> typeof <: ZeroMeanDiagNormal for the checks.
Later can dispatch on this to a chainrules callback if it is not this type.
Relies on #33
MvNormal(Diagonal([0.1, 0.2])) |> typeof <: ZeroMeanDiagNormalfor the checks.Later can dispatch on this to a chainrules callback if it is not this type.