Mabwiser

- Bug fixes in examples
- Validate tree parameters of TreeBandit to be compatible with sklearn.tree.DecisionTreeRegressor

- Implement vectorized functions for non-contextual policies to speed-up prediction for multiple decisions.
- Change MAB predict and predict_expectations to allow empty contexts to be specified for non-contextual policies.
- Update scaler use in Linear policies so that standard scaler can be fit directly instead of pre-trained scalers.
- Change scaler argument from pre-trained `arm_to_scaler` input to a boolean scale flag.

- New Algorithm: LinGreedy as a learning policy.

- Modified `predict_expectations`, such that `predict` can use `predict_expectations` directly in all non-contextual learning policies.

- Added warm_start method to MAB, that allows untrained (cold) arms to be warm started based on features of each arm.
- Added remove_arm method to MAB to allow arms to be removed from bandit.

* Breaking: Updated NumPy RNG backend to utilize the new Generator class. This is a breaking change for algorithms with random components.
* Updated NumPy version dependency to >=1.17.0 to reflect the utilization of the new Generator class.
* Updated multivariate sampling logic in LinTS to utilize updated NumPy RNG backend