Pick-and-place Manipulation Across Grippers Without Retraining: A Learning-optimization Diffusion Policy Approach