Stable Recovery of Entangled Weights: Towards Robust Identification of Deep Neural Networks from Minimal Samples