Better support for `:=`

[torfjelde]

There are a few things missing from the := to reach peak utility:

• Tracked variables are only included in calls to sample => not currently used in returned or predict, i.e. it's non-trivial to extract := values using a chain.
• := values are included in the chain returned from sample. However, it is treated as a standard parameter in the resulting chain! This means that it's not possible for downstream functionality, e.g. predict, to determine which parameters are actual parameters taht should go into, say, a VarInfo, and parameters that are from := expressions and thus should not go in, say, a VarInfo. Ref: Track := variables in resulting chains #693
• Some things we'd like to track involve expensive computations that are unnecessary for inference. Hence we should ideally have a way to only activate this when we're not sampling. Ref: Convenience macros to use within @model #714 Docs on sub-model workflow for generated quantities, predictions, and performance optimisations #810 (comment) is preferred as an alternative workflow instead of introducing more macros.