NNUE 2: Augment net input features with classical eval score

[xefoci7612]

The idea is to add classical eval score to the NNUE input feature set.

The aim is to let the net, not only to have a second important source of information on the position, but to learn, when this information is valuable and when is not and hence to automatically train to output a value that is always the best between the 2 worlds, even better than choosing between classical and NNUE values, because possibly will be a learned smart blend between the two. We don't need anymore "use classical" stuff.

If the impact on the Stockfish speed is too big, then a middle ground could be used:

1. Train a teacher net with the feature set augmented by classical evaluation score

2. Train a standard NNUE student net to mimic the teacher and hence indirectly get some understanding of classical knowledge, still avoiding to call classical evaluation at every move.