This integrates the Scalable Checkpoint / Restart library into the DeepSpeed checkpoint path. The major changes are:

• add start / complete calls to declare the start and end of each checkpoint phase,
• avoid creating any checkpoint directories when using SCR (SCR will create those when needed),
• rely on SCR to record and report the tag name for a restart,
• register each checkpoint file and acquire a temporary path to use when writing the file.