Add particle detection features

[kephale]

Overview

This PR adds a comprehensive set of particle detection features to the copick-torch library, making it easier to detect particles in cryoET data:

1. MONAI-based Particle Detector

• Implements a 3D object detector based on RetinaNet from MONAI
• Supports both 2D and 3D inputs
• Includes training and inference pipelines
• Sliding window inference for large tomograms
• Customizable anchors and detection parameters

2. Difference of Gaussian (DoG) Detector

• Classic blob detection algorithm reimplemented for particle picking
• Optimized for cryoET data
• Parameter optimization against ground truth
• Multiscale detection capabilities
• Simple but effective implementation

3. Evaluation Metrics

• Precision, recall, F1 calculation with tolerance-based matching
• Average Precision (AP) calculation
• Helper functions for distance calculation
• Comprehensive detector evaluation tools

4. CryoET Data Portal Integration

• DataPortalDataset for loading tomograms with automatic resolution rescaling
• ParticleDataset for training with automatic background sampling
• Efficient caching mechanisms
• Integration with the copick and cryoet_data_portal libraries

Example Scripts

• Full example for the DoG detector
• Command-line interface for training and inference with the MONAI detector

Tests

• Comprehensive test suite for all components
• Automated tests for all core functionality

Testing

All components have been tested with unit tests.

Impact

This PR provides easy-to-use implementations of particle detection algorithms that will help users analyze their cryoET data more effectively.

Future Work

• Add more particle detector types
• Improve training capabilities with larger datasets
• Optimize inference speed further