[9/9] local_main_process_first when building dataset

[garrett361]

Changes the main_process_first ctx manager wrapping data creation to local_main_process_first, which is more appropriate for multi-node use cases (and works for single-node as well).

	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]

Converted to draft, due to concerns that local_main_process_first seems like it can cause errors in multi-node scenarios where all nodes share the same HF cache. When this is the case, it appears there may be race conditions where multiple procs attempt to delete the same cache files, leading to errors.