Add `prob_sub` argument to `init` of algorithms

[Red-Portal]

If we want to use DifferentiationInterface.prepare_gradient with the parameter space algorithms, we need to call prepare_gradient for differentiating through the objective. Now, prepare_gradient assumes that the types don't change. Currently, the "unconditioned" prob is passed on to AdvancedVI.init so that prepare_gradient can be called. When subsampling is involved, however, the type of the target problem changes from the "full objective" prob to the minibatch-conditioned objective prob_sub. Therefore, prepare_gradient can fail. This needs to be fixed by potentially adding an argument to step.

[Red-Portal]

After considering the details, this is unnecessary.