PipelineTrainer KL step lag

[Kovbo]

Summary

• Rebase the KL step-lag behavior from Replace kl_penalty_reference_step with kl_penalty_step_lag #625 onto the current main branch without changing the contract.
  • kl_penalty_coef=0: KL disabled
  • kl_penalty_coef>0 and kl_penalty_step_lag=None: use checkpoint 0
  • kl_penalty_coef>0 and kl_penalty_step_lag>=1: use max(0, current_step - lag)
  • kl_penalty_step_lag=0: invalid