Specify loss as a part of distribution_output for torch models

[karthickgopalswamy]

Right now we have to specify distr_output and loss to instantiate estimator object, but the choice of loss cannot be arbitrary as its connected to the type of distr_output (e.g. StudentT or Gaussian Output need NegativeLogLikelihood but QauntileOutput such as ISQF need CPRS). As in the example, since its tied for a choice of Output why don't we add loss to DistributionOutput object with a get_loss method?

@jaheba @lostella