When training a model, all steps in the forward pass are tracked to compute gradients of loss in the backward pass. Please read about GradientTape if you are interested in learning more about this.
Using the final lambda layer and multiplying by a constant is equivalent to dividing all features by the constant before feeding to the input layer. Read this link to understand the importance of feature scaling.
As far as the constants 100 and 400 are concerned, the staff must’ve picked these hyperparameters with experimentation on model loss with different scaling values. My recommendation is to always perform feature scaling before feeding data to the model.