Lost-ds

Changed
- Moved package building to pyproject.toml
Fixed
- Fixed Pipy build

Fixed
- Try to fix pypy build

Changed
- `split_train_test` now filters samples with too few unique sources to prevent errors in sklearn
- refactored to_abs and to_rel method for easier readability and better performance
Added
- retry reading a parquet file without opening as buffer if it fails
- semantic_segmentation can store files with numeric names
- new function: `split_train_test_multilabel` for fairly splitting multilabel datasets
- added requirements
Fixed
- vis_and_store can handle multilabel, singlelabel and None correctly now
- polygon validation fixed. Set required amount of points from 4 to 3.
- fixed in `validate_unique_annos` pandas drop() future warning
- `crop_dataset` does not produce empty duplicates of crop positions anymore
- `segmentation_to_lost` fix bug where some contours of different classes were merged accidentially
- `mask_dataset` won't overwrite input dataframe anymore
- added a serialization case where a columns has empty lists

Added
- fixed voc_eval bug where gt_df bboxes weren't handled correctly
- fixed coco_eval bug - different dataframes having same uids for images and labels now
- to_coco meethod has arguments predef_img_mapping and predef_lbl_mapping now to hand over predefined label and image uids
- export method crop_img to lost_ds (lost_ds.crop_img and ds.crop_img formerly lost_ds.DSCropper.crop_img)
- resolve FutureWarning at transform_bbox_style
- make crop_dataset compatible with pandas > 1.5.0
- crop_component indexing is imagewise instead of global
- improve some filter and transform functions
Breaking changes
- Add parallel option for multiple functions
- fix typo in some methods arguments

Added
- method `split_train_test` which allows the stratify option for i.e. classification datasets
- `coco_eval` method for detection (mAP, Average Recall)
- `voc_eval` method for detection (tp, fp, fn, precision, revall, ap)
- `to_coco` improvements and bugfixes
- `voc_score_iou_multiplex` method: shifting bbox score and iou_thresholds to find optimal thresolds
- cropping method return additional column 'crop_position' now
- added arg 'random_state' for dataset-splitting
- added arg for optional parallelistaion
- added color selection for vis and store - can take column now
- improved detetion metrics

Changed
- file_man: Allow other Fsspec filesystems
Added
- Progress callback for pack_ds
- Argument `fontscale` at `vis_and_store` to enable manually control for textsize
- Argument `cast_others` at `segmentation_to_lost` to allow ignoring unspecified pixel values