[peft] Add support for LoRA merge

[ananthsub]

Offer API to merge base model weights with LoRA adapters given a path to a LoRA checkpoint generated from megatron bridge.

1. Use the serialized ConfigContainer in the run_config.yaml to read the pretrained_checkpoint field from the CheckpointConfig - this points to the path for the base model checkpoint directory.
2. We re-use the existing functionality from model_load_save for initializing the base model from the serialized config and loading the base weights.
3. Then use the instantiated LoRA transform from the run_config.yaml and apply the transform.
4. After the transform is applied, we load the adapter weights from the provided LoRA checkpoint path.
5. Once the adapter weights are loaded, the LoRA Merge transform is applied.
6. Finally, we generate the config metadata for the merged checkpoint, and save the weights for the merged model.

Parity with https://github.com/NVIDIA/NeMo/blob/42bc4b20b170cb65ed15e4691f9e769ca2941865/nemo/collections/llm/peft/api.py#L76-L80

[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.

[ananthsub]

/ok to test d8ded7c

[ananthsub]

/ok to test 71d4668

[ananthsub]

/ok to test 34f6564

[marcromeyn]

/ok to test 38d0ed2

[marcromeyn]

/ok to test ba5deb5

[github-actions]

This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.