RAFL: Generalizable Sim-to-Real of Soft Robots with Residual Acceleration Field Learning