Add general gradient estimator to SampledObs class.

[markusschmitt]

For an expectation value

$\langle O\rangle_\theta=\sum_s\frac{|\psi_\theta(s)|}{\langle\psi_\theta|\psi_\theta\rangle}O_\theta(s)$

the general form of the gradient is

$\partial_\theta\langle O\rangle_\theta=2\langle\text{Re}[\Gamma_\theta] O_\theta\rangle_\theta-2\langle\text{Re}[\Gamma_\theta]\rangle_\theta \langle O_\theta\rangle_\theta+\langle\partial_\theta O_\theta\rangle_\theta$

where $\Gamma_\theta(s)=\partial_\theta\log\psi_\theta(s)$.

This PR implements a mean_and_grad member function of the SampledObs class that computes the gradient of the expectation value as given above.