To be or not to be stable, that is the question: understanding neural networks for inverse problems