[6/9] sync_each_batch option for grad accumulation

[garrett361]

Adds a sync_each_batch argument for optionally synchronizing gradients after every batch when using gradient accumulation. Specifying --sync_each_batch True can dramatically reduce memory costs when using FSDP/DeepSpeed. The default is False, preserving the upstream open-instruct behavior.

	PR	Title
1	#15	padding-free
2	#16	clean_checkpoints_at_end
3	#17	final_lr_ratio
4	#18	add_seed_and_date_to_run_name
5	#19	additional_model_arguments
6	>20	sync_each_batch=True grad acc
7	#21	no grad acc averaging for sum losses
8	#22	extra reporting
9	#23	local_main_process_first when building dataset

[garrett361]

I'm going to make this configurable with the previous behavior as the default