Support parallel SGMCMC without needing `torch.vmap`

[SamDuffield]

Currently the user can run many independent chains of sgmcmc by utilizing torch.vmap themselves. However, torch.vmap doesn't compose with a lot of existing pytorch code. It might be nice to allow a boolean flag like batched or batched_param or parallel that allows the user to define the log_posterior that takes on input and output an additional batch dimension. This way parallel sgmcmc could be run without needed torch.vmap