Importing NVRx to call the corresponding routines migrated to NVRx

[sbak5]

Is your feature request related to a problem? Please describe.
Adding dependency in MCore to call corresponding routines in NVRx for checkpointing.

We're in the middle of migrating the checkpointing at dist_checkpointing.

Tag the @mcore-oncall
to get oncall's attention to this issue.

Describe the solution you'd like
CI pipeline / any checkpointing routines having dependency on the migrated routines will run only when NVRx is installed. If not, it will run corresponding torch.distributed.checkpoint routines.

Async ckpt will be enabled only with NVRx installed.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

[sbak5]

Three steps:

1. Import NVRx to run the same routines as we have now and see if there's no bug or difference between importing the routines from NVRx and using the routines in MCore. (changes in core/dist_checkpointing/strategies/torch.py, training/async_utils.py)
2. Prepare a path without NVRx for synchronous checkpointing. (changes in serialization.py, base.py)
3. Update the pyproject.toml / uv.lock to bind the dependency in MCore to a release or commit in NVRx.

[sbak5]

@sbak5 shared a pip wheel with @dimapihtar to initiate these steps

[sbak5]

https://github.com/sbak5/Megatron-LM/tree/sbak/ckpt_migrate
includes changes for step 1.
@dimapihtar please refer to this and work on others on top of this. (review and refinement is required though)