Generalize compute_gradient

[jrmaddison]

Just as a tangent-linear computes a general directional derivative (a JAX jvp) the adjoint computes a (conjugate) derivative of a general dual element evaluation (the conjugate of a JAX vjp).

For a state $u \in V$ the generalization is defined by a functional

$J (u) = \eta (u)$,

for some $\eta \in V^*$, and the adjoint then computes (the conjugate of) the derivative of the reduced functional $\hat{J} = J \circ \hat{u}$ where $\hat{u}$ defines the forward (mapping from the control to the state).

compute_gradient is currently the case where $u$ is a scalar-valued output and $\eta$ is the identity.

This is already possible using InnerProduct or DotProduct to define the functional, and this is already used for the Gauss-Newton approximation. A simple interface could be added, and a new marker might be added to allow $\eta$ to be changed without re-evaluating the forward.

[jrmaddison]

AdjointActionMarker introduced in #542 could be used for this.