Text this: Cost-Efficient Distributed Learning via Combinatorial Multi-Armed Bandits