Smlb

This release adds the `results` given to a `LearningCurvePlot` to the evaluation's auxiliary data. (This makes it easy to reference the data passed to the evaluation after the fact.)

This release updates dependencies, adds an optimization trajectory example and and fixes a few bugs in example notebooks. We also dropped support for Python 3.6 and added support for Python 3.9.

**What's New**

* We added a new optimization trajectory example.

**Improvements**

* Library dependencies were updated to their latest versions.

**Fixes**

* We fixed a few errors in the example notebooks.

This release adds more functionality to optimization trajectory workflows and makes imports simpler.

**What's New**

* The `IdentityLearner`, which mimics a synthetic function. This allows for the use of a synthetic dataset as a surface on which to test optimization (15).
* "Rook design", a simple optimizer that allows you to investigate how step size and pruning impact the ability of an optimizer to work on a given response surface (15, 25).
* Package-level imports. Now all featurizers can be imported from `smlb.features`, all learners can be imported from `smlb.learners`, all optimizers can be imported from `smlb.optimizers`, and all workflows can be imported from `smlb.workflows`. Datasets are divided into experimental/synthetic, and can be imported from `smlb.datasets.experimental` or `smlb.datasets.synthetic` (21, 25).

**Improvements**

* Optimization trajectory results are now plotted as a line with shading instead of as a set of scatter plots. The median score over all trials is drawn as a solid line, and specified quantiles are shaded. The default is to shade from the 0.25 to the 0.75 quantile. The max/min values can optionally be drawn as dashed lines (22).
* The `ExpectedValue` score can now specify that lower values are optimal (23).

**Fixes**

* Fixed a bug that caused optimizers to drive towards the wrong target when used with `ExpectedValue` scores (20).

This is the first official release of the Scientific Machine Learning Benchmark (smlb) package. smlb is focused on enabling rigorous empirical assessments of data-driven modeling approaches for applications in the natural sciences. This release includes the following:

* a variety of synthetic and experimental datasets
* CDK and Matminer featurizers
* random forest model backed by [Lolo](https://github.com/CitrineInformatics/lolo) as well as a variety of random forest and Gaussian process models backed by [Scipy](https://www.scipy.org/)
* standard loss functions and error metrics, including those for evaluating the quality of predictive uncertainty
* reproducibility by systematic control of pseudo-random number generation
* two global optimizers backed by [Scipy](https://www.scipy.org/)
* two types of workflows: one to compare model quality with learning curves and the other to compare the efficacy of different optimizers

For more details, see the [overview](https://github.com/CitrineInformatics/smlb/blob/master/docs/overview.md). Example Jupyter notebooks can be found [here](https://github.com/CitrineInformatics/smlb/tree/master/examples).

Testing