Add more docs on benchmarking

[penelopeysm]

Right now our only mention of TuringBenchmarking is in the AD bit

[willtebbutt]

Would you consider expanding the scope of this issue to a more general issue on analysing performance problems, and fixing them? e.g. how do I figure out what bit of my model is slow, using standard Julia profiling tools? What do you, as the user, have control over?