- Introduced a graph reordering mechanism to improve computational performance
  on GPUs.
- The default starting probability of the marginalized graph kernel is now
  hyperparameterized and will be optimized by default during training.
- Allow users to specify custom starting probability distributions.
- Performance improvements due to the in situ computation of starting
  probabilities instead of loading from memory.
- Added `repeat`, `theta_jitter` and `tol` options to the Gaussian process
  regressor.
- Fixed a normalization bug in `GaussianProcessRegressor.fit_loocv`.