Leveraging a residual learning framework to support the model training on non-ideal analog in-memory computing hardware