Egobox

* `gp`:
* Add analytic derivatives computations (54, 55, 56, 58, 60). All derivatives available for all mean/correlation models are implemented.
* Refactor `MeanModel` and `CorrelationModel` methods:
* `apply()` renamed to `value()`
* `jac()` renamed to `jacobian()`
* Fix prediction computation when using linear regression (52)
* `ego`:
* Refactor `Egor` using [`argmin 0.7.0` solver framework](http://argmin-rs.org) `EgorSolver` can be used with `argmin::Executor` and benefit from observers and checkpointing features (#67)
* `Egor` use kriging setting by default (i.e. one cluster with constant mean and squared exponential correlation model)
* Add [notebook on Manuau Loa CO2 example](https://github.com/relf/egobox/blob/master/doc/Gpx_MaunaLoaCO2.ipynb) to show `GpMix`/`Gpx` surrogate model usage (#62)
* Use xoshiro instead of isaac random generator (63)
* Upgrade `ndarray 0.15`, `linfa 0.6.1`, `PyO3 0.17` (57, 64)

* `gp`: Kriging derivatives predictions are implemented (44, 45), derivatives for Gp with linear regression are implemented (47)
* `predict_derivatives`: prediction of the output derivatives y wtr the input x
* `predict_variance_derivatives`: prediction of the derivatives of the output variance wrt the input x
* `moe`: as for `gp`, derivatives methods for smooth and hard predictions are implemented (46)
* `ego`: when available derivatives are used to optimize the infill criterion with slsqp (44)
* `egobox` Python binding: add `GpMix`/`Gpx` in Python `egobox` module, the Python binding of `egobox-moe::Moe` (31)

* Add Egor `minimize` interruption capability (Ctrl+C) from Python (30)
* Minor performance improvement in moe clustering (29)
* Improvements following JOSS submission review (34, 36, 38, 39, 40, 42)

* Generate Python `egobox` module for Linux (20)
* Improve `Egor` robustness by adding LHS optimization (21)
* Improve `moe` with automatic number of clusters determination (22)
* Use `linfa 0.6.0` making BLAS dependency optional (23)
* Improve `Egor` by implementing automatic reclustering every 10-points addition (25)
* Fix `Egor` parallel infill strategy (qEI): bad objectives and constraints gp models updste (26)

Improve mixture of experts (15)

* Implement moe save/load (feature persistent)
* Rename GpSurrogate to Surrogate
* Remove `fit_for_predict`
* Implement `ParamGuard` for `MoeParams`
* Implement `Fit` for `MoeParams`
* Rename `MoeParams` setters

Refactor `moe`/`ego` relation (16)

* Move `MoeFit` as `SurrogateBuilder` from `moe` to `ego`
* Implement `SurrogateBuilder` for `Moe`
* `Moe` uses `linfa::Fit` trait
* Rename `Evaluator` as `PreProcessor`

Refactor `MixintEgor` (17)

* Rename `PreProcessor::eval` to `run`
* Implement `linfa::Fit` for `MixintMoeParams`, use `linfa::Dataset`
* Rename `SurrogateParams` to `MoeBuilder`
* Rename `n_parallel` to `q_parallel` (qEI stategy)

* Improve documentation
* `egobox` Python module: rename egobox `Optimizer` class to `Egor`