Introduce optimizer_base_type in support of different optimizers

[milancurcic]

This is the first step toward decoupling the optimizer logic from the concrete layers.

This PR only introduces the abstract optimizer_base_type and a concrete sgd type.

The update of weights is still hardcoded in network % train and the concrete layer implementations; decoupling that remains a TODO.

In a nutshell, the idea is to have

• Concrete optimizer types such as sgd, adam, etc. in nf_optimizers.f90 (and its submodules, eventually);
• The type constructors would expect optimizer parameters from the user (e.g. adam(learning_rate, beta1, beta2, epsilon, ...))
• Each concrete type would define an update subroutine which would expect the needed gradients (dw, db) as input, and also the weights and biases arrays as intent(out) to update.

@rweed let me know if this approach seems reasonable to you.