Deepspeed-chat feature request: specify which epoch/iteration to start EMA

leveraging learnings from the mosaic ml blogpost on speeding up diffusion model training https://www.mosaicml.com/blog/diffusion it doesn't seem like we need to update the EMA model weights at every iteration of training, the last <5% of steps should be fine.

i propose adding an argument like --ema_start_pct with a default like 0.95, so start_ema_iter = int(args.ema_start_pct *num_total_iters), then keeping a global track of current_iter` such that

if args.enable_ema and current_iter >= start_ema_iter:
    moving_average(rlhf_engine.actor,
                                rlhf_engine.actor_ema,
                                zero_stage=args.actor_zero_stage)

Start in the DeepSpeed-chat training entry point by locating argument parsing, the total-iteration calculation, current_iter tracking, and the moving_average call on rlhf_engine.actor and actor_ema. Add the requested EMA start percentage behavior and verify that EMA updates begin only at the configured iteration, including the default value.