Extract the per-level search from `speed_hierarchical()`

[rogerssam]

speed_hierarchical() (R/speed.R) has grown to the point where its per-level
body is around 130 lines, and each feature added to it makes the next extraction
harder. feature/keep-with added three more concerns: swappable-group discovery,
the unequal-replication warning, and stop_reason bookkeeping.

What the function is actually doing is a fold over levels. The body of that fold
is a self-contained simulated-annealing run, and it should be a helper:

.optimise_level <- function(design, opt, level, grid_idx, quiet, ...) {
  # random_initialise has already run; this is one level's SA loop
  return(list(
    design         = best_design,
    scores         = scores,
    temperatures   = temperatures,
    stop_reason    = stop_reason,
    optimal_score  = optimal_score,
    best_score_obj = best_score_obj
  ))
}

That would leave speed_hierarchical() as the loop over hierarchy_levels
plus the metadata assembly it already does in a second pass, and would make the
per-level state (scores, temperatures, n_kept, last_improvement_iter,
temp) local to the run that owns it rather than reassigned in an outer scope.

Worth doing as its own change rather than inside a feature branch - it is a pure
refactor with no behavioural component, and the test suite is a good net for it.

Not urgent. Filing so it is not rediscovered on every review of this area.