[BUG]: ZeRO checkpoint would change the offload behavior

[Gy-Lu]

🐛 Describe the bug

In the current code, there is a lack of processing for offload when comes to checkpoints.
For example, with offload, the device of optimizer state should be CPU, which leads to an error when comm.

Besides, when loading the checkpoint, the device of the optimizer state should follow the original one rather than the loaded one. However, now it's just the opposite.

Environment

No response