Consolidate retraining epochs into per-fold JSON and silence parallel fold logs

[Copilot]

The parallel CV runner in examples/quantum/classification.py wrote a redundant epochs.json even though each fold_<n>_results.json already carried the retraining history, and worker processes printed heavily interleaved progress messages to the terminal.

Changes

• Drop epochs.json: removed the post-run aggregation block. Per-epoch history stays where it already lives — inside each fold's own checkpoint (q_epochs, q_final_epoch, q_final_error).
• Quiet parallel workers: run_fold now executes its body under contextlib.redirect_stdout(os.devnull), so concurrent folds no longer race for stdout. The main process still renders the stderr progress bar during the run and prints the aggregated CV summary once all folds complete.

# examples/quantum/classification.py
def run_fold(fold_num, ...):
    checkpoint_file = f"fold_{fold_num}_results.json"
    with open(os.devnull, "w") as _devnull, contextlib.redirect_stdout(_devnull):
        return _run_fold_body(fold_num, ..., checkpoint_file)

The fold body is factored into _run_fold_body purely so the redirect wrapper doesn't force re-indenting the whole function; behavior of the fold computation and checkpoint format are otherwise unchanged.