Text this: Visual–Tactile Fusion and SAC-Based Learning for Robot Peg-in-Hole Assembly in Uncertain Environments