Pytorch-segmentation-models-trainer

- Minor improvements and bug fixes on polygon building inference;
- Bug fixes on mask builder;
- Performance improvement on mask builder using coco format;

- Added inference features;
- Improved polygon inference;

- Changed the versions of pytorch and torchvision.

- Added MANIFEST.in to include missing yml on pypi packaging.

- Bug fix on loss sync;
- Custom models from Frame Field implementation (to compare training results);
- New HRNet-OCR-W48 backbone;
- Fixed bugs on new versions of pytorch-lightning;
- Build mask from COCO dataset format;

- Polygon inference
- Unittests to Polygon inference;
- Bug fixes warmup callback (invalid signature on method);
- FrameFieldResultCallback renamed to FrameFieldOverlayedResultCallback;
- New implementation of FrameFieldResultCallback;
- Invalid mask handling (frame field training mask with only polygon mask and empty vertex and boundary masks);
- Added multiple schedulers option;
- Added IoU 10, 25, 50, 75 and 90;
- Added GPU augmentation using kornia;