Add `expect_and_forces` method to variational states

Created by: femtobit

Gradients of observables are estimated in VMC via the covariance F_j = \mathrm{cov}(O_j, E_\mathrm{loc}) where O_j = \partial\ln\psi(s)/\partial\theta_j is the derivative of the log-probability. This is called the force vector here.

For real parameters, the observable gradient is then \partial\langle \hat O\rangle / \partial\theta_j = 2\mathrm{Re}[F_j]. For a complex holomorphic parametrization \partial\langle \hat O\rangle / \partial\theta_j^* = F_j.

Note that F is generally complex even for real-parameter models. Since expect_and_grad returns the gradient and thus only the real part of F, information is lost in this step, which is needed in particular in time evolution (where the RHS of the equation of motion given by McLachlan's variational principle involves \mathrm{Im}[F_j]).

expect_and_forces adds a way to get the original forces for use in TDVP. I've added some tests and docs, the specific implementation in this PR has been written by @PhilipVinc.