Expected value iterations

[odow]

Talking to Gabriel: we should add an option to run some expected values cuts (off by default).

This might be very useful in cyclic problems to cheaply get the value function to a place where it is the right height before adding more cuts.

We could do this by averaging the noise terms. We probably need an API like:

SDDP.train(
    model;
    average_value_iterations = SDDP.AverageValueIterator(; num_iterations = 50) do P, Ω
        return sum(p * ω for (p, ω) in zip(P, Ω))
    end,
)

[odow]

So perhaps the right thing to do here is an expected value forward pass. Or just pick the middle realization from the vector. It doesn't need to be the expected value...