Migrate from custom runner and MONAI ConfigParser to Lighter

[ibro45]

Hey guys, I've just played around with this now that Lighter's been redesigned to be simpler and more Pytorch-Lightning-native. Actually, the redesign was partially prompted by this repo. Not expecting this to be merged, just for your interest.

Summary

• Replace MONAI's ConfigParser with Lighter framework for configuration management
• Add __lighter__.py marker file enabling project.* imports automatically
• Update all YAML configs to Lighter syntax (%vars::, $@ expressions)
• Rename config sections: lightning_module → model, data_module → data
• Update Python imports to use project.* prefix throughout codebase
• Simplify dependencies in pyproject.toml
• Remove scripts/run.py (replaced by lighter CLI)

New Usage

# Training
lighter fit configs/train.yaml configs/models/primus.yaml configs/datasets/amos.yaml

# With CLI overrides
lighter fit configs/train.yaml trainer::max_epochs=50 model::base_lr=0.0005

# Prediction
lighter predict configs/predict.yaml

Benefits

• Cleaner CLI: lighter fit config.yaml instead of custom fire scripts
• Better override syntax: model::learning_rate=0.01
• Automatic timestamped experiment outputs
• Built-in config saving for reproducibility

Test plan

• Verify lighter CLI is available after install
• Test config parsing with trainer::fast_dev_run=true
• Full training run with actual dataset