This fixes the mishandling of in_progress and the queues in the submit functions that are in part irq driven, too, and ensures nothing breaks by preventing concurrent access from irqs to the queues. Applying the same pattern to all functions/cases is fine either way.
On top of all of tha the out ep was irq driven, so if submit_next_out fails it was stuck, now fixed by attempted submit from main loop, and of course the queue manipulation should be irq safe here as well.
