Use step_warmup for AdaptiveHamiltonian

[penelopeysm]

Closes #2678

Note, CI not running because this needs TuringLang/AbstractMCMC.jl#178 to go through.

This PR:

• Removes the nadapts argument from HMCDA and NUTS constructors. Instead of specifying that, the user should specify the num_warmup keyword argument to sample().
• Removes the discard_adapt keyword argument when sampling with HMCDA/NUTS, in favour of discard_initial, which has the same meaning for all samplers and is therefore more coherent.
• Since adaptive Hamiltonians must do different behaviours for adaptation / warmup steps, this PR also implements AbstractMCMC.step_warmup for these samplers.

This also needs:

• Changelog

[penelopeysm]

Oh, for goodness' sake, this is way more annoying than I had realised. DynamicPPL.Sampler, simply, straight up, does NOT allow for step_warmup at all because DynamicPPL.initialstep is only called via step and not step_warmup,

[penelopeysm]

I think this will just have to wait for Sampler to be removed as a whole.