Catasta

new features
- the `Archway` class has seen a complete overhaul, now supporting multiple ways of inference: standard forward, onnxruntime, or a fastapi endpoint for remote inference
- the model in the `Archway` can be compiled and quantized using TorchScript before inference

breaking changes
- removed `PredictionInfo` dataclass

- fixed a bug where it just didn't work because of a broken import :P

new features
- keyboard interrupt stops training instead of raising an exception
- the examples under `examples` numbered and easier to follow !!!!!
- added a `batch_size` argument to the `Scaffold.evaluate` method
- the models can be exported to the `onnx` format
- catasta now supports multi-output regression
changes
- console logging is now more consistent (and more beautiful imo)
- the `path` argument of the `Scaffold.save` method can also be used with a filename

breaking changes
- the `max_lr` argument in the `Scaffold.train` method has been removed. the schedulers will return in a future update with more flexibility
- removed accuracy-related metrics in the `TrainInfo` class
- renamed `ApproximateGPRegressor` to just `GPRegressor`
- changed the argument name `context_length` of regressors to `n_inputs`
- changed the name of classifiers to `[Model]ImageClassifier` to support signal classification in the future
- changed the name of `CatstaDataset` to just `Dataset`
- removed `GPFormerRegressor`, `PatchGPRegressor`, and `PatchGPFFTRegressor` to include the `GPHeadRegressor`

New features
* A new class, called `Foundation`, has been added to optimize the hyperparameters of the models. Check the `examples` for more information.
* Added a new module `catasta.utils` with the following new util functions:
* `split_dataset` helps splitting datasets into training, validation, and testing folders.
* `set_seed` let's you set a seed for reproducibility
* `set_deterministic` sets PyTorch flags for reproducibility
* Added the following new models:
* PatchGPRegressor
* PatchGPFFTRegressor
* GPFormerRegressor

* The pooling of the `TransformerRegressor` and `TransformerFFTRegressor` can be selected.
Breaking changes
* Removed the possibility to select the `dtype` of the model in the `Scaffold`.
* Changed the `early_stopping` argument in `Scaffold.train` from a boolean value to the smooth factor `alpha`.
* Changed the `min_lr` argument in `Scaffold.train` to `max_lr` as it uses `OneCycleLRScheduler` now.

Minor changes
* Performance should have increased a bit
* CUDA cache is cleaned before training
* An error will raise if a wrong loss function is chosen

New Features
* The images now can be loaded as grayscale.
* Added an argument to the train method of the `Scaffold` to choose the number of workers of the data loader.
Changes
* The training message logging has been prettified.
* In the `ApproximateGPRegressor`, the `use_ard` option defaults now to True.
* The `path` argument in the `Archway` is no longer a keyword argument.
* Changed the type of the `task` argument in the `CatastaDataset` from `Literal` to `str`.
* General performance has been improved.

Buf fixes
* Disabling verbose did not hide all training messages.

Breaking Changes
* `ClassificationEvalInfo` and `RegressionEvalInfo` have been unified into a single class called `EvalInfo`. To differentiate them, the task will be provided via a string, as in the `CatastaDataset `.
* The attribute `input_size` in the `FeedforawrdClassifier` has been changed to `imput_shape`.
* Deleted MAPE, SMAPE, MASE, and MASEP metrics.
* Simplified the `CNNClassifier` constructor arguments.

New features
* Improved training logging by displaying multiple rows instead of a single one.

Bug fixes
* Now, only if the loss function is a `MarginalLogLikelihood`, the loss will be negated.
* Fixed a bug where images were not loading for classification if the extension was in uppercase.
* The `CatastaDataset` now accepts a list of strings in the argument `input_name` to select multiple inputs for regression.