Many miscellaneous fixes; the Plato framework has also passed the type checker using Astral's ty.

Supported MLX as a new backend (an alternative to PyTorch), and added Torchvision as a new datasource, retiring MNIST, EMNIST, CIFAR10, CIFAR100, FashionMNIST, and CelebA.

Migrated the configuration file format across-the-board from YAML to TOML.

Completed the migration of the client, server, and trainer APIs to the use of the strategy design pattern for better separation and composition.

Incorporated LoRA fine-tuning and improved the Hugging Face trainer.

Added support for secure multi-party computation.

This release redesigned the client, trainer, and server APIs to use a strategy-based composition approach.

Redesigned the trainer API to use a composable trainer architecture, based on strategy patterns and callbacks, rather than subclassing and hooks, which is more error-prone.

Migrated the entire framework to be compatible with uv, PyTorch 2.8, and Python 3.13. Removed support for MindSpore and TensorFlow.

Added support for self-supervised learning and personalized federated learning.