Fast and Sample Efficient Multi-Task Representation Learning in Stochastic Contextual Bandits