Sotai

Bug Fixes:
- updated API response object keys to be snake_case instead of camelCase so that retrieval works properly.
- updated analysis URL to properly point to overall-model-results tab in the client so that the generated URL is no longer broken.

api:
- uploading a dataset (`post_dataset(...)`)
- launching a hypertune job (`post_hypertune_job(...)`)
- retrieving a cloud hosted pipeline (`get_pipeline(...)`)
- retrieving the list of trained model UUIDs for a given cloud hosted pipeline (`get_trained_model_uuids(...)`)
- retrieving cloud hosted trained model metadata (`get_trained_model_metadata(...)`)
- downloading a cloud hosted trained model (`download_trained_model(...)`)

enums:
- HypertuneMethod enum for defining the method to use for hypertuning. Currently limited to grid search.

layers:
- `assert_constraint(...)` functions for each layer (CategoricalCalibrator, NumericalCalibrator, Linear) for asserting that specified constraints are satisfied.

models:
- `assert_constraints(...)` function for `CalibratedLinear` model to assert satisfaction of constraints for all contained layers.

pipeline:
- trained_models are now stored in a dictionary `trained_models: Dict[int, TrainedModel]` that maps the ID (not UUID) to the `TrainedModel` instance.
- `hypertune(...)` method for launching a hyperparameter optimization job either locally or on the SOTAI cloud for the pipeline.
- pipeline saving will now include trained models by default in the save path under a `trained_models/{id}` folder structure. Any trained models present in this structure will be loaded. If you'd like to exclude models from loading, simply store them elsewhere. You can also choose to exclude trained models from saving and save them manually yourself (i.e. to save only the models you want to use again and drop all other models).
- `from_hosted(...)` method for loading a cloud hosted pipeline.

trained_model:
- `TrainedModel` is now a subclass that extends the new `TrainedModelMetadata` type.
- `from_hosted(...)` method for loading a cloud hosted trained model.

types:
- `HypertuneConfig` type for defining how to hyperparameter optimize your models during a hypertune job.
- `TrainedModelMetadata` type container for all trained model metadata excluding the model itself.

api:
- functions for uploading a trained model, running inference, getting inference job status, and getting completed inference results.

pipeline:
- `analysis(...)` now uploads the trained model to the SOTAI web-app.
- `inference(...)` function for running inference with a given trained model.
- `await_inference(...)` function for awaiting and then downloading inference results into a specified file.
- `from_config(...) class method for initializing a pipeline from a `PipelineConfig`.

fixes:
- projection_iterations is now configurable from pipeline -> model.
- `FeatureAnalysis` can now have optional statistics. This is necessary for categorical features that do not have such statistics.
- categories can be initialized using integer values without getting auto-converted into strings. This prevents a bug where the categories are strings but the underlying data is integers.

api:
- functions for creating a pipeline, pipeline config, feature configs, and trained model on the SOTAI web-app through the API.

pipeline:
- analysis function for analyzing a trained model in the SOTAI web-app.

Initial SOTAI SDK Beta Release

pipeline:
- Pipeline object for defining how to ingest data and train calibrated models.
- TrainedModel object that contains the trained model, configurations, and results from running `pipeline.train(...)`

types:
- Defines the Pydantic types used throughout the SDK. Most of these objects act as containers or configs.

models:
- CalibratedLinear model for easy construction of a calibrated linear model using feature objects.

features:
- NumericalFeature and CategoricalFeature objects used for easy creation of calibrated models.

layers:
- NumericalCalibrator for calibrating numerical features. Supports bound and monotonicity constraints.
- CategoricalCalibrator for calibrating categorical features. Supports bound and pairwise monotonicity constraints.
- Linear for linearly combining inputs. Supports monotonicity and weighted average constraints.

data:
- CSVData class for loading, preparing, and batching data for calibrated modeling.
- General helper functions for replacing values and identifying feature types.

training:
- Training utilities. Generally used for the implementation of Pipeline.

enums:
- String Enums used throughout the SDK. Any Enum typed value can be set with the string equivalent.

constants:
- Constants used throughout the SDK.