Specifying both directional derivative and gradient

[vincentmolin]

In some applications the gradient vector product dot(∇ϕx, θ) is significantly cheaper to compute than the full gradient ∇ϕx. For the samplers where it makes sense to support this, it could be nice to implement it.

(Would be pretty cool if ∇ϕx could be lazy in some sense, dot(∇ϕx,v) evaluated the directional derivative and reflect!(∇ϕx,... forced a full gradient evaluation.)