Egobox

* `ego`:
* Allow to reuse surrogate trainings (reuse previous hyperparameters) from a previous iteration by relf in <https://github.com/relf/egobox/pull/157>
* Fix hot start mechanism due to bad loaded DOE exploitation by relf in <https://github.com/relf/egobox/pull/156>
* Adjust log messages by relf in <https://github.com/relf/egobox/pull/158>
* Log Egor configuration by relf in <https://github.com/relf/egobox/pull/159>
* `gp`:
* Allow fixed hyperparameters theta for GP and Sparse GP by relf in <https://github.com/relf/egobox/pull/155>
* `egobox`:
* Upgrade to pyo3 0.21 by relf in <https://github.com/relf/egobox/pull/152>
* Upgrade dependencies by relf in <https://github.com/relf/egobox/pull/153>

* Fix GP mixture with kpls option on Griewank test function by relf in <https://github.com/relf/egobox/pull/150>

* [Breaking changes] `gp`, `moe`, `egobox` (Python): Rename `predict_derivatives()` as `predict_gradients()` by relf in <https://github.com/relf/egobox/pull/148>

* [Breaking changes] `gp` API renaming by relf in <https://github.com/relf/egobox/pull/145>
* `predict_values()` is renamed `predict()`
* `predict_variances()` is renamed `predict_var()`
* `predict_variance_derivatives()` is renamed `predict_var_derivatives()`
Derivatives predictions (`predict_derivatives()` and `predict_var_derivatives()`) are made available in Python.
* Refactor Mixture of Experts by relf in <https://github.com/relf/egobox/pull/146>
Factorize code between full GP and sparse GP implementations

* Add `Gpx` accessors by relf in <https://github.com/relf/egobox/pull/140>
* Fix `LHS` maximin bug by relf in <https://github.com/relf/egobox/pull/141>
* `doe`: Improve classic, centered and maximin LHS performances by relf in <https://github.com/relf/egobox/pull/138>
* `doe`: Improve optimized LHS performances (1.25x speedup) by relf in <https://github.com/relf/egobox/pull/136>
* Rework (mostly internal) API to avoid awkward &Option by relf in <https://github.com/relf/egobox/pull/134>
* Add Python bindings for all LHS flavours by relf in <https://github.com/relf/egobox/pull/135>

* `gp`: Implement sparse gaussian process methods (cf. `SparseGaussianProcess`)
* Python binding: `SparseGpMix`, see doc/tutorial
* GP/SGP API
* hyperparameter tuning : initial theta guess and bounds can be specified (`theta_init`, `theta_bounds`)
* `n_start` controls the number of optimization multistart
* In GP/SGP `rayon` is used to make parallel optimization multistart