Load schedulers from ckpt in a backward-compatible manner

[hyperzy]

• The loading behavior is similar to that in sdfstudio.
• Allow loading old checkpoints without the schedulers key.
• Versions <= 0.3.2 don't support loading schedulers state_dict, because the states of schedulers are not saved in Trainer.save_checkpoint(). Specifically, commits older than cebe787 don't support schedulers loading/saving.
  These changes in the commit will save the states of schedulers by default, i.e. creating a schedulers key. Were the checkpoints from previous versions loaded, the loader will check whether the schedulers key exist before loading schedulers.

[tancik]

LGTM