Support parallel argument to predictive

[fehiepsi]

In Pyro MCMC predictive, we have a parallel keyword to resolve the plate issue pyro-ppl/pyro#1995. JAX supports batching out of the box. However,
using vmap for a large num_samples and a large model will consume much more memory. I faced memory issue in training bnaf for covtype dataset and using lax.map solves the problem.

Probably we should change the keyword to sequential to avoid confusion.

[neerajprad]

That's a good idea, we should have this.