Slisemap

Additions:
- Add procedures for hyperparameter optimisation (see `slisemap.tuning`).
- Added wrappers that combine `local_model`, `local_loss`, and `coefficients` into one name (see, e.g., `slisemap.local_models.LogisticRegression`).

Changes:
- Change defaults for `slisemap.metrics.accuracy` (to `optimise=False`, `local_loss=True`).
- Add `numpy` parameter to `Slisemap.fit_new`.
- Improve the `Slisemap.predict` function to take more parameters.
- Move `entropy` to `slisemap.metrics`.
- Fix some deficiencies in `slisemap.utils.Metadata`.
- Improve the x-axis labels for the matrix plot.

New features:
- Metadata
- Add metadata about e.g. variable names and normalisation
- The metadata is reused for all plots
- More general input
- Anything that can be turned into a `torch.Tensor` or `numpy.ndarray` is accepted
- This includes dataframe-like objects (also imports variable names)

Improvements:
- Better colourscales for local losses
- Centering before PCA
- `Slisemap.load` understands `map_location` from `torch.load`
- More verbosity levels for optimisation

Changes:
- Remove types from docstrings
- Rename "Fidelity" to "Local loss" in plots

Deprecations:
- With the addition of metadata, some parameters to the plotting functions have been deprecated

What's Changed
* Refactor demo examples + add binder demo to README by MomoLangenstein in https://github.com/edahelsinki/slisemap/pull/3
* Point the user towards multiple_linear_regression if they use nd-y by MomoLangenstein in https://github.com/edahelsinki/slisemap/pull/5
* Improve docs by Aggrathon in https://github.com/edahelsinki/slisemap/pull/6
* Cleanup of properties (some deprecation warnings)
* Update references by Aggrathon in https://github.com/edahelsinki/slisemap/pull/7


**Full Changelog**: https://github.com/edahelsinki/slisemap/compare/v1.1.0...v1.2.1

Main changes in this release:

- Improved documentation.
- Less randomness (only perturb the embedding if a loss of rank is detected).
- New argument `only_B` in the `Slisemap.optimise` (the same as `Slisemap.lbfgs(only_B)`).
- Tweaks to the plots and updated experiments.
- A notebook discussing how to use PyTorch for optimisation.

- Exposed `get_model_clusters`.
- Made `scikit-learn` a required dependency (and removed all optional dependencies).
- Fixed a bug when calculating PCA using SVD.

- Fix a bug with the scaling in `fit_new`
- More verbosity in `fit_new` and `optimise`
- Add logistic regression with log-loss to `local_models`