Token Pooling-Aware Training for Multi-Vector Embedding Models