Robust estimation in generalized linear models: the density power divergence approach