Efficient Training of Deep Neural Operator Networks via Randomized Sampling