ModPhase-like model for ARNNDense

Created by: VolodyaCO

I have been minimising infidelities for some spin problems using the ARNNDense model with complex parameters, but the optimisation is always stuck in some excited state. If I rewrite the ARNN to separate the modulus and the phase (like the RBMModPhase), the problem disappears and I'm able to correctly minimise the infidelity.

If you think that these ModPhase-like models are useful, I can write a class that should be like an AbstractARNNModPhase, and have the conv and dense ARNNs inherit from it. Please let me know what you think.