[Feature]: Evaluate semi-automatic perception dataset bootstrapping with Roboflow workflows

[ll7]

Feature Description

PAF already has a local DVC-managed traffic-light dataset and an Ultralytics-based perception stack, but expanding or refreshing datasets still looks expensive to do fully by hand.

The most promising Roboflow contribution here is not their hosted runtime stack. It is the workflow ideas shown in their public notebooks around:

• Grounding DINO / SAM style auto-annotation
• fast bootstrap labeling for new object-detection or segmentation datasets
• YOLO fine-tuning loops on exported datasets

This issue is to evaluate whether a local, repo-compatible version of that workflow can help PAF create or extend datasets for traffic lights and other perception classes without coupling the repository to Roboflow cloud features.

Definition of Done

• Select one or two candidate workflows from the public Roboflow examples that fit PAF best.
• Run a small trial on an existing PAF dataset slice or a newly generated CARLA sample.
• Measure annotation quality, cleanup effort, and export compatibility with the current training code.
• Decide whether the resulting workflow is worth adopting for traffic lights, general object detection, or neither.
• Record the outcome and any guardrails, especially that datasets should remain local and DVC-managed unless there is a strong reason to change that.

Testability

Use a representative labeled subset and compare the semi-automatic workflow against the current manual process in terms of time spent, correction effort, and downstream usefulness for training or evaluation.

Dependencies

Related to #918, since better CARLA-side dataset capture would create stronger inputs for this work. This issue can still start earlier on the existing traffic-light dataset or other small local samples.