MAINT/FIX: enable transfer/continuous learning with initialized net

[nghi-truyen]

• Fix initialize method for net: only initialize weight and bias if it is None, enabling transfer/continuous learning. This feature is already possible in the previous version since net.compile method was previously separated. In this version, it can be used as follows:

  model = smash.Model(*smash.factory.load_dataset("lez"))
  ret = model.optimize("ann", 
                    optimize_options={"termination_crit": dict(epochs=10), "random_state":11}, 
                    return_options={"net": True})  # pre-train the neural network
  net = ret.net.copy()  # net is now the pre-trained model
  net.set_trainable([0,0,0,0,1,0,0])  # set trainable layers
  ret2 = model.optimize("ann", 
                    optimize_options={"net": net, "termination_crit": dict(epochs=50)}, 
                    return_options={"net": True})  # continuous learning

• Fix set_trainable method: do not update weight and bias when trainable is set to False.