feat: MFU logging in train recipes

[SwekeR-463]

What does this PR do ?

• Adds AutoMFU to compute and log Model FLOPs Utilization (MFU) in LLM training recipes, with registry-style model/device handling under _transformers.

Changelog

• Added new AutoMFU class in _transformers.
• Added registry-style device mapping for MFU reference TFLOPs.
• Exported AutoMFU from _transformers/__init__.py.
• Integrated MFU reporting into LLM train recipes in train_ft.py and train_seq_cls.py.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?

Additional Information

• Related to MFU logging in train recipes #1385

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[hemildesai]

Can you add something like https://github.com/verl-project/verl/blob/main/verl/utils/flops_counter.py#L22-L85 for the device registry?

[SwekeR-463]

Can you add something like https://github.com/verl-project/verl/blob/main/verl/utils/flops_counter.py#L22-L85 for the device registry?

Have done the required change.

[akoumpa]

/ok to test 7e43c4b

[akoumpa]

/ok to test 3bc3ccc

[akoumpa]

/ok to test 32bd853