Use vectorize in `finite_discrete_marginal_rv_logp`

[ricardoV94]

This should be more performant and also faster to compile due to smaller computational graphs.

I've been waiting for the vectorize functionality to mature, and I think it's good now