Multi-task Representation Learning for Pure Exploration in Bilinear Bandits