Kinisi

Changes
- SciPy `1.7.0` is now required on installation

Changes
- The variance is now determined from using generalised least squares on a number of random samples from the Gaussian process, this has the benefit of being more accurate and efficient.
- As a result of the above there is no longer a "Sampling Likelihood" `tqdm` bar.
- There is now a `rtol` kwarg for the `bootstap_GLS` method, this controls the `rtol` in the call to the `scipy.linalg.pinvh` that is used to stabilise the large matrices.
- All documentation has been adapted to reflect this methodological change.

Changes
- Add the ability to use multiple identical simulations from an `MDAnalysis.Universe` class object.

Changes
- Add the `flatchain` property to the `Analyzer` classes, this returns the samples from the MCMC sampling of the diffusion parameter and intercept.
- Include the use of the new `flatchain` property in the [documentation](https://kinisi.readthedocs.io/en/latest/vasp_d.html).
- Pin to the latest (`1.2.5`) version of `uravu`, which includes the `flatchain` property for `Relationship` class objects.
- Use the `flatchain` property from `uravu` in the [Arrhenius tutorial](https://kinisi.readthedocs.io/en/latest/arrhenius_t.html).

Changes
- Significant improvements to documentation, including a detailed description of the [methodology](https://kinisi.readthedocs.io/en/latest/methodology.html) used by `kinisi`, a tutorial showing the use of the [Arrhenius](https://kinisi.readthedocs.io/en/latest/arrhenius_t.html) functionality and the addition of a [covariance matrix derivation document](https://kinisi.readthedocs.io/en/latest/_static/derivation.pdf).
- Running and testing on python 3.10 (via CI and support in setup.py).
- A transition in the API to follow a more `pymatgen` style, where `Analyzer` objects are accessed by calling a `from_*` method based on the inputs.
- Addition of a logo file to the repository.

Changes
- Previously, the sampling of the diffusion coefficient allowed for negative values, this was determined to be unphysical. Therefore, a Bayesian prior of a [Heaviside step function](https://en.wikipedia.org/wiki/Heaviside_step_function) at 0 has been assigned to the gradient/diffusion coefficient to stop this.
- It was noted that for very large number of atoms/timesteps (which are typically handled using `MDAnalysis`), there could be a memory overflow. This is due to the creation of a list of numpy arrays, for which the largest could be in the gigabyte range.

item_size * atoms * timesteps * dimensions * to Gb
8 * 20000 * 10000 * 3 * 1e-9 = 4.8 Gb

Therefore, in addition to being able to `sub_sample_traj`, it is also possible to `sub_sample_atoms` (current this is only supported of `MDAnalysis` objects), which performed the same function as `sub_sample_traj` but on the atoms instead of timesteps.
- The code is now fully [typed](https://docs.python.org/3/library/typing.html) and formatted to follow PEP8 formatted guidelines (with the exception of a 120 max line length).