Recovering Plasticity of Neural Networks via Soft Weight Rescaling