Fix #1064: Auto log sample train/val images at training start (with tests)

[bw4sz]

Description

This PR implements automatic logging of sample images from training and validation datasets at the start of training, as requested in #1064.

Changes

Implementation (src/deepforest/main.py)

• Added on_train_start hook that:
  • Samples up to 5 images from training dataset
  • Samples up to 5 images from validation dataset (if available)
  • Uses visualize.plot_annotations() to create annotated images
  • Logs images to all available experiment loggers (Comet, TensorBoard, W&B, etc.)

Tests (tests/test_on_train_start.py)

• Comprehensive test suite including:
  • Test that training images are logged correctly
  • Test that validation images are logged correctly
  • Test with multiple loggers
  • Test with empty annotations (edge case)
  • Test that correct number of images are sampled
  • Test without any loggers (should not crash)
  • Test that visualize.plot_annotations is called correctly
  • Test with no validation dataset
  • Test that parent class behavior is preserved

Implementation Details

• The hook checks for non-empty annotations before sampling
• Images are saved to a temporary directory
• Each image is logged with metadata including filename, context (train/val), and current step
• Compatible with all loggers that have a log_image method
• Gracefully handles edge cases (empty datasets, no logger, etc.)

Testing

Run the new tests with:

pytest tests/test_on_train_start.py

The implementation will automatically log images when training starts. Users will see sample images in their experiment tracking tool of choice.

Fixes #1064

[jveitchmichaelis]

This is the one from Cursor? It does have some "tells" - for some reason Claude seems to really like mocking in unit tests (and it loves to make long test suites). I had to make a rule to discourage that.

Can the bot respond to PR comments on GitHub? It would be great if we could feedback/review issues and have it fix stuff automatically.

[bw4sz]

Waiting on #1079 to see if there are rebase conflicts.

[henrykironde]

Looks good but the tests could use some changes

[bw4sz]

@henrykironde this should be ready, passes locally.