Refactor MNIST example and enhance training options

- Updated the MNIST GPU example to include TrainOptions for better training configuration.
- Simplified the loading of MNIST data and ensured proper handling of existing files.
- Improved the neural network architecture in the MNIST example for better performance.
- Added a new drawing example for MNIST digit classification using eframe and egui.
- Moved serialization functionality for saving and loading neural network models.
- Enhanced the training module with a callback system for monitoring training progress.
- Cleaned up the backend module to ensure only one backend feature can be active at a time.
- Fixed minor issues in the transpose shader comments and ensured consistent formatting.

Author: NLion74

.gitignore

@@ -1,3 +1,4 @@
 /target
 *.bin
-**/data/**
+**/data/**
+examples/testing