Performance hints

[torfjelde]

We should have decent ways to inform the user of certain simple things they can do to improve the performance of their model.

For example, in #2396, it is suggested that we should maybe warn a user if they use a forward-mode AD backend on a model with more than, say, 20 parameters. This is ofc a nuanced issue, as not all AD backends support the same (sub)set of Julia.

However, we could maybe hide some of these suggestions behind a function, e.g. performance_hints(model) or something, which prints out certain suggestions to how to improve a model / warns the user of typical patterns that have performance impacts.

Again, it'll be difficult to make this perfect, but if accompanied with good docs (ref: #2415), this could work well.

Example of things we could put in such a performnace_hints method:

1. Usage of reverse-mode AD backend when dimensionality is "high".
2. Using DynamicPPL.check_model_and_trace we could also analyse the ~ statements to check if there's anything going on there, e.g. x[i] ~ Normal() in a loop instead of x ~ filldist(Normal(), n).
3. More?

Longer term, we could probably make use of JET.jl to inspect more things, e.g. are allocations happening in the model, etc.