Mixed precision accumulation for neural network inference guided by componentwise forward error analysis