`AbstractMCMC.step` interface broken about advancedhmc.jl

[devmotion]

Just remembered TuringLang/AbstractMCMC.jl#124.

I still think this should be fixed in AdvancedHMC rather than AbstractMCMC - we can't add all downstream options to AbstractMCMC and it means that custom implementations based on step (as in my use case) will always use a default value of 0, if they don't know about and reimplement the logic in AdvancedHMC.

[yebai]

@devmotion @torfjelde @JaimeRZP do we have a default mechanism for users to pass warmup strategy somewhere? If not, it would be good to introduce an AbstractWarmUp type in AbstractMCMC so we have a consistent interface for passing warmup strategies in AbstractMCMC.step functions.

[devmotion]

TuringLang/AbstractMCMC.jl#117 will allow samplers to specify behaviour of warm-up steps. AdvancedHMC would specialize step_warmup, possibly depending on a warm-up algorithm instance stored in the sampler.