Add PipelineTrainer checkpoint artifact uploads

[arcticfly]

Summary

• add a save_checkpoint_artifact parameter alongside save_checkpoint on PipelineTrainer
• upload saved eval checkpoints to W&B as LoRA artifacts via the existing W&B deployment helper
• normalize compatible Unsloth Llama base-model aliases to W&B-supported metadata IDs
• add focused unit coverage for the new trainer parameter and W&B base-model aliasing

Validation

• uv run ruff check src/art/pipeline_trainer/trainer.py src/art/utils/deployment/wandb.py tests/unit/test_pipeline_trainer_local_backend.py
• python -m py_compile src/art/pipeline_trainer/trainer.py src/art/utils/deployment/wandb.py
• import/signature check confirmed PipelineTrainer.__init__ includes save_checkpoint_artifact
• downstream Willow smoke job saved W&B LoRA artifacts at step1 and step2, then both artifacts responded to W&B inference requests

Note: the full ART unit file collection was blocked locally by an existing environment mismatch: transformers requires huggingface-hub<1.0 but the ART venv has huggingface-hub==1.7.2.