Learning Mixed Multinomial Logits with Provable Guarantees