Summary:
Adds an experimental implementation for a rank local checkpointer with save and load with partial load, blind load and in-place load.
This uses an new API and simpler format.
Plan to add async checkpointing, IO layer, pluggable storage backend, layout customization, Resharding, deduplication etc are not implemented.
Test Plan: unit tests
Reviewed By: saumishr
Differential Revision: D75426560
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156142
Approved by: https://github.com/saumishr