Describe the workflow you want to enable

I would like SGDRegressor to be able to accept additional values in its 'loss' arguments to be able to incrementally train additional regressors, that are currently available in SkLearn in the form of non-incremental regressors (PoissonRegressor / GammaRegressor/TweedieRegressor).

Describe your proposed solution

Add additional loss function implementations (providing gradients) to support the additional regressor types.

Describe alternatives you've considered, if relevant

Learning with L1/L2 loss in log-space. But it doesn't work well when the label of a given sample is zero. Poisson regression handles it gracefully.

Additional context

No response