Add ancestor resampling option for PG

[yebai]

The PG sampler can be improved by adding the ancestor resampling step. It is challenging to implement this for the general case, but perhaps we can consider adding it a a option, e.g.

s = PG(10, 20, :s, as = true)

Here when as equals true, we support only one variable (e.g. :s) to be sampled by PG.

References:

• http://jmlr.org/papers/volume15/lindsten14a/lindsten14a.pdf
• https://papers.nips.cc/paper/5968-particle-gibbs-for-infinite-hidden-markov-models.pdf

[emilemathieu]

Insights from Jan-Willem van de Meent at TuringLang/Turing.jl#334 (comment)

[donhausk]

I am currently working on ancestor sampling. If someone else is also working on this, please contact me to have a short discussion.

[yebai]

Closed by #32