wip: add detr keypoint architecture

[jveitchmichaelis]

@bw4sz no direct training integration yet, but this is what I'm thinking for the modeling.

The implementation looks verbose, but that's because a lot of the pieces are copied from transformers with very minor changes - like a reference to bbox changed to point. To modify the arch we need:

DeformableDetrKeypointConfig: Same as object detection but explicit parameter for point loss cost.
DeformableDetrKeypointDetectionOutput: Mostly similar but name changes.
DeformableDetrForKeypointDetection: Model, changes number of decoder outputs to 2 and removes some bbox specific bits.
DeformableDetrKeypointMatcher: Hungarian matcher that uses L1 or L2 cost instead of IoU.
DeformableDetrKeypointLoss: L1 or L2.
DeformableDetrKeypointImageProcessor: Normalizes keypoints to relative pixel coords and some other bits.

I don't anticipate checking in the current test suite as the overfit test takes a minute or two to run, but it does converge.

[jveitchmichaelis]

High level:

1. Dataset support, initially derived from existing box datasets where we can derive keypoints from box centers.
2. Configuration task support which is tied to architecture, e.g. task = boxes by default (allows retinanet, detr), but can be task = points (detr), polygons (?)
3. Modeling support (create/load models should open the appropriate architecture)
4. Transforms + augmentation should be minimal for now, but ensure that they support keypoints if passed in
5. Training support - ensure that forward pass works with .fit() and all functions in main work
6. Metrics should be modified to be task-specific
7. Evaluate should also allow keypoint precision/recall based on matching metrics, with a threshold for pixel distance instead of IoU

[codecov]

Codecov Report

❌ Patch coverage is 78.57143% with 138 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.52%. Comparing base (915a945) to head (4ed7f25).
⚠️ Report is 23 commits behind head on main.

Files with missing lines	Patch %	Lines
src/deepforest/models/keypoint.py	79.48%	64 Missing ⚠️
src/deepforest/datasets/training.py	71.42%	22 Missing ⚠️
src/deepforest/main.py	72.60%	20 Missing ⚠️
src/deepforest/evaluate.py	77.61%	15 Missing ⚠️
src/deepforest/predict.py	73.33%	8 Missing ⚠️
src/deepforest/utilities.py	80.95%	4 Missing ⚠️
src/deepforest/augmentations.py	57.14%	3 Missing ⚠️
src/deepforest/callbacks.py	75.00%	1 Missing ⚠️
src/deepforest/visualize.py	75.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1182      +/-   ##
==========================================
- Coverage   87.38%   85.52%   -1.86%     
==========================================
  Files          20       22       +2     
  Lines        2569     3165     +596     
==========================================
+ Hits         2245     2707     +462     
- Misses        324      458     +134     

Flag	Coverage Δ
unittests	85.52% <78.57%> (-1.86%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[jveitchmichaelis]

@bw4sz integration is done, I'll add tests. Lots of TODOs dotted around which relate to tidying up and merging some duplicated code, but I anticipate polygons/segmentation will follow in a similar way.

The pre-commit failure is weird. It thinks that pandas is undefined, but the import is unchanged and ruff passes locally.

[jveitchmichaelis]

Currently training a DETR backbone on the lidar pretrain dataset, will swap that in once done.

[jveitchmichaelis]

Will remove this, or make it an option.