Rectools-lightfm

Fixed

- Re-Cythonized cython files to fix compilation errors with newer compilers.
- Fixed `np.object` usage in tests.

Addded
- Set the `LIGHTFM_NO_CFLAGS` environment variable when building LightFM to prevent it from setting
`-ffast-math` or `-march=native` compiler flags.

Changed
- `predict` now returns float32 predictions.

Added
- Added a check that there is no overlap between test and train in `predict_ranks` (thanks to [artdgn](https://github.com/artdgn)).
- Added dataset builder functionality.
Fixed
- Fixed error message when item features have the wrong dimensions.
- Predict now checks for overflow in inputs to predict.
- WARP fitting is now numerically stable when there are very few items to
draw negative samples from (< max_sampled).

Added
- added additional input checks for non-normal inputs (NaNs, infinites) for features
- added additional input checks for non-normal inputs (NaNs, infinites) for interactions
- cross validation module with dataset splitting utilities
Changed
- LightFM model now raises a ValueError (instead of assertion) when the number of supplied
features exceeds the number of estimated feature embeddings.
- Warn and delete downloaded file when Movielens download is corrputed. This happens in the wild
cofuses users terribly.

Added
- added get_{user/item}_representations functions to facilitate extracting the latent representations out of the model.
Fixed
- recall_at_k and precision_at_k now work correctly at k=1 (thanks to Zank Bennett).
- Moved Movielens data to data release to prevent grouplens server flakiness from affecting users.
- Fix segfault when trying to predict from a model that has not been fitted.

Changed
- Ranks are now computed pessimistically: when two items are tied, the positive item is assumed to have higher rank. This will lead to zero precision scores for models that predict all zeros, for example.
- The model will raise a ValueError if, during fitting, any of the parameters become non-finite (NaN or +/- infinity).
- Added mid-epoch regularization when a lot of regularization is used. This reduces the likelihood of numerical instability at high regularization rates.