Awkward thinning/burn-in semantics for `sample_chain`

Currently, sample_chain implements thinning and burn-in as follows:

• the first sample will step the inner kernel num_burnin_steps + 1 times
• any other samples will step num_steps_between_results + 1 times

This runs into the issue of never accomplishing thinning with repeated calls to the driver with num_steps=1. To illustrate, consider the toy example:

current_state = 0
for _ in range(5):
  new_sample, _ = sample_chain(
      num_steps=1,
      current_state=current_state,
      kernel=<some_kernel>,
      num_burnin_steps=0,
      num_steps_between_results=100)
  current_state = do_something(new_sample)

At every iteration, the first sample drawn from the kernel is returned from sample_chain, though one should expect some thinning. The workaround would be doing everything manually from kernel.one_step calls, but that kind of defeats the purpose of having a smart driver in the first place.

It seems more natural to have semantics as follows:

• the first sample will step the inner kernel num_burnin_steps + num_steps_between_results + 1 times
• any other samples will step num_steps_between_results + 1 times

Start at the sample_chain entry point and reproduce the repeated num_steps=1 example with num_steps_between_results=100. Done means the first returned sample includes burn-in plus thinning steps, while subsequent samples retain the stated thinning interval.