Functional Output Regression with Infimal Convolution: Exploring the Huber and $\epsilon$-insensitive Losses