Multi-domain training, integrated gradients and API updates

[jacobkimmel]

• Adds multi-domain training support with a new scnym_api argument: domain_groupby
• Replaces raw gradient magnitude-based saliency methods with integrated gradients
• Adds implementations of baseline models benchmarked in the scNym manuscript
• Adds hyperparameter tuning through reverse cross-validation
• Updates README to highlight multi-domain training
• Removes pre-trained models to avoid improper use
• Updates tests