The dataset fails to load with `datasets==3.6.0` and `pyarrow==20.0.0` due to a type conversion error:
```
pyarrow.lib.ArrowInvalid: Float value 0.000012 was truncated converting to int64
```
This occurs because the JSON data contains floating-point values in bounding box coordinates, but the dataset schema expects `int64` values.
## Solution
- Added `convert_bbox_to_int()` helper function to properly convert float coordinates to integers using `int(round(coord))`
- Applied conversion to both `bboxes_block` and `bboxes_line` in `_generate_examples()`
Tested with:
- `datasets==3.6.0`
- `pyarrow==20.0.0`
- Successfully loads all splits (train/validation/test) without errors
