Speed up backward message for `Linearization`

[bartvanerp]

The computation of the backward message for the Linearization approximation around the Delta node can be sped up significantly, without the need of computing the marginals over all inputs. This can be achieved by using more effective computations in the forward pass and weightedmean precision parameterization in the backward pass.

[bvdmitri]

Related https://juliadiff.org/ForwardDiff.jl/stable/user/advanced/