VLA-Touch: Enhancing Vision-Language-Action Models with Dual-Level Tactile Feedback