feat: Resume functionality for online data mixing

[kmehant]

Changes

Online data mixing dataset (ODM) is wrapped as stateful dataloader along with internal individual datasets. The load and store of the state dicts are defined accordingly to fully save and load the state of the ODM dataset. Since transformers do not support stateful datasets yet, we do patching of related functions to allow usage of ODM as stateful dataloader seamlessly. Dataloader state is saved as "odm_dl_state_dict.bin" file in individual checkpoint folders.

Resumption graphs (using fms-hf-tuning)

Settings:

param	value
eval_steps	3
save_steps	2
per_device_batch_size	8
FSDP distributed ranks	2
update_interval	every 3 steps
sample_interval	every batch
reward	validation_loss

param	value
eval_steps	3
save_steps	2
per_device_batch_size	8
FSDP distributed ranks	2
update_interval	every 3 steps
sample_interval	every batch
reward	entropy