Use `step_warmup` for AdaptiveHamiltonian rather than `nadapts`

[penelopeysm]

NUTS currently has a parameter called nadapts which controls how many MCMC steps include an adaptation bit.

discard_adapt then controls whether those adaptation steps are discarded at the start.

For consistency with the AbstractMCMC API this should be changed to be num_warmup + discard_initial. And the adaptation steps should call AbstractMCMC.step_warmup rather than AbstractMCMC.step.

[penelopeysm]

Note that this is blocked by DynamicPPL, as its implementation of initialstep is only called by AbstractMCMC.step and not AbstractMCMC.step_warmup. Thus if we use step_warmup together with DynamicPPL.Sampler we lose anything that is called inside initialstep. Consequently this cannot be fixed until we remove Sampler which would in turn allow for more customised behaviour of Turing samplers (e.g. using step_warmup properly).

This is done.

[penelopeysm]

See TuringLang/AbstractMCMC.jl#203 for some design questions.