Avoid CPU OOM when loading full-state FSDP checkpoints in Fabric

[awaelchli]

What does this PR do?

Fixes #18008
Related: #8043

Resolves a comment in the code regarding the loading of full-state checkpoints into a FSDP model: Before this PR, each worker loads its own copy of the checkpoint into CPU memory. Example: On a machine with 8 GPUs and a checkpoint of size 10 GB, we would occupy 8 * 10 = 80 GB of CPU memory when loading the state dict. If there is not enough CPU RAM, we get OOM.

This PR implements a strategy in which we load the state dict sequentially among the local ranks. The drawback is that loading will take N times longer, where N is the number of GPUs on the machine.

• Was this discussed/agreed via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

Reviewer checklist

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified