Pedestal-inference

- Added `numba` as an additional option for JIT compilation. Previously the `jit_compile` keyword argument expected a `bool`, but now accepts either `"jax"`, `"numba"` or `None` to select what jit-compilation is used.
- Added tools to support easier performance benchmarking.

- Increased minimum required Python version from `3.7` to `3.9`.
- The method for calculating predictions of spectrum measurements has been re-designed in order to improve numerical efficiency and enable jit-compilation via [jax](https://github.com/google/jax). As a result the computation time of the posterior log-probability has been reduced by roughly a factor of 5.
- jit-compilation via `jax` can be enabled by setting `jit_compile=True` when creating an instance of `SpectralResponse`.
- `jax` has been added as an optional dependency, which can be included during installation using `pip install pedestal-inference[jit]`.

Change to profile model interface
In order to allow for more complex profile models (e.g. those which include configurable basis functions) the interface to models has had to change slightly. Methods of the model, e.g. `mtanh.prediction`, are no longer static, and so can not be called directly from the class. For example, you could previously do
Python
from pedinf.models import mtanh
fit = mtanh.prediction(radius, parameters) this will now raise an error

Now instead an *instance* of the model class needs to be created:
Python
from pedinf.models import mtanh
model = mtanh() create an instance of the model class
fit = model.prediction(radius, parameters)


Other changes
- Renamed the `SpectralPedestalPosterior` class to `ThomsonProfilePosterior` as it better describes its function.
- Added a new profile model `exspline` which is is appropriate for fitting the profile from one edge into the core. It combines a chosen set of exponentiated b-splines to model the background profile with a logistic function to model the pedestal structure.

- Optimised the `SpectrometerModel.spectrum` function.
- Added analytic calculation of the Jacobian of the `SpectrometerModel` predictions to allow for gradient-based optimisation of the log-posterior.

- Fixed a bug where `SpectrumData` would not properly sanitise it's input data if all the data were `NaN`.
- Added the ability to pass a prior function to `SpectralPedestalPosterior`

- Fixed a bug in `SpectrometerModel` which caused a crash when accessing the shape of the spectral response data array.