Inference-tools

- Added a new class `GpLinearInverter` for performing Gaussian-process linear inversion.
- Added a new covariance function class `ChangePoint`.
- Hyper-parameter labels for mean and covariance functions have been redesigned to be more specific.
- `GpRegressor` has a new keyword argument `n_starts` which allows the number of BFGS starting positions used during hyper-parameter optimization to be set manually.

- Added a `display_progress` keyword argument for MCMC classes which can be used to suppress progress message prints during sampling.
- Added an additional plot-style option to `inference.plotting.matrix_plot` which draws contours using highest-density intervals.

- Added a new MCMC sampling class `EnsembleSampler`, which is an implementation of the 'affine-invariant ensemble sampler'.

- Improved the efficiency of linear algebra calculations in `GpRegressor` related to hyper-parameter optimisation.
- Initial guess generation for acquisition function maximisation in `GpOptimiser` now works properly in cases where some of the data are outside the search bounds.
- The choice of optimiser used for acquisition function maximisation in `GpOptimiser` can now also be specified when calling `GpOptimiser.propose_evaluation`.

- Added a `WhiteNoise` covariance function to model the presence of Gaussian noise on input data for Gaussian-process regression.
- 'Composite' covariance functions can now be constructed via addition - see the updated [GP regression notebook demo](https://github.com/C-bowman/inference-tools/blob/master/demos/gp_regression_demo.ipynb).
- The `gp_tools` and `pdf_tools` modules have been renamed to `gp` and `pdf` respectively. This is not a breaking change - importing using the old module names is still supported, but now raises a warning.

Fixed an import in `inference.__init__` which was causing issues for python 3.6 and 3.7 in some cases.