Change user-facing random variable API

[ricardoV94]

This PR gets rid of RandomStream, and provides as an alternative random methods on RandomGeneratorType variables.

So one can do:

rng = random_generator_type("rng")  # symbolic equivalent to np.random.Generator
next_rng, x = rng.normal()
fn = function([rng], x)
fn(np.random.default_rng(1)  # draws

rng can obviously be a shared variable, and have an update {rng: next_rng}, which is a bit what RandomStream is doing under the hood.

However, everything is now explicit. There are no longer any default_updates, and unlike numpy the user is given the output next_rng to use afterwards. This is a bit awkward, but it's the most explicit API?

The RandomStream had important issues in rewrites, because of the default_update graphs are implicit, and when one mutates a graph with such variables they would also need to mutate the default_update graph for the final thing to be correct.

RandomVariable Ops will no longer hide the next_rng output like they used to. There's a transition period forcing you to commit to the new flag return_next_rng, which will eventually default to True if unspecified

next_rng, x = pt.random.normal(rng=rng, return_next_rng=True)