label-studio-converter

## Bug Description:
When executing the command **label-studio-converter import yolo -i data -o ls-tasks.json** to generate a JSON file for pre-annotated YOLO detection images, the following folder architecture is assumed:
data:
- images
- labels
- classes.txt
The images folder contains **5537** images, and the labels folder contains an equal number of txt files corresponding to annotations. However, the generated JSON file only contains **3948** images. I've confirmed that each image has a corresponding txt file, and they were generated using a trained YOLOv8 model.

## Environment:
Label Studio latest Docker image

Start by reproducing `label-studio-converter import yolo -i data -o ls-tasks.json` with the described images, labels, and classes.txt structure, then inspect how the converter discovers and serializes YOLO inputs. Done means the generated JSON includes all 5537 images with their corresponding annotations.

Written by the indexing model from the issue text.