Refnx

------
- Drop Python 3.6 for new development (as per
https://numpy.org/neps/nep-0029-deprecation_policy.html)
- Pin setuptools version on RTD to allow the docs to build.
- Add a tutorial to the docs on model selection.
- Update emcee vendored code.
- Fix repr of Model
- Fix several warnings during test run.
- Add error bar plotting to refnx GUI.
- Improved the specular ridge finding methodology for data reduction.

------
- build more Linux wheels.
- vendored ptemcee (bleeding edge), so it's no longer necessary to install it.
- use -funsafe-math-optimizations for reflectivity calculation.
- fixed bug when retrieving log-probabilities from PTMCMC sampling.
- fixed encoding call for Py3 (.decode doesn't exist for Py3).
- improve pymc3 model creation. pymc3 offers different ways to carry out MCMC
sampling. This may lead to pymc3 being included as a sampling method in
Curvefitter.
- implement `invcdf` method for Bounds objects. This will allow creation of
prior transforms from [0, 1) to the original range specification.
- BREAKING CHANGE: `objective.model.logp` and `Objective.logp_extra` are now
included in the log-likelihood (`Objective.logl`) instead of log-prior
(`Objective.logp`). This makes it easier to work with pymc3 and dynesty.
- Add `Objective.prior_transform` that converts random variates in a unit
hypercube to parameter values, according to their prior distributions.
- manual_beam_find can be sent a name to be displayed on the window title.
- ReductionOptions dict is now used for specifying options for reduction and
processing of reflectometry data. This will permit different settings to be
used for datasets measured in different configurations.
- Add AutoReducer, an object that watches a directory for modified/created
files and automatically reduces them.

------
- Build standalone GUI as part of continuous integration runs.
- Fix couple of GUI warnings when started from a terminal.
- pyinstaller plist settings adjusted to make retina compatible and include
refnx version number.
- use Github Actions to test and build Linux wheels across all the Pythons.
- Compensate for dq_type being added to ReflectModel (older .mtft files
wouldn't load into GUI).
- Refactor test_reflect to challenge all reflectivity calculation backends.
- refnx.reflect.available_backends lists all reflectivity calculation backends.
- Add pytest fixture to download test data. This dramatically reduces the
size of the package.
- Compensate for _stderr being added to Parameter (older .mtft files wouldn't
load properly into GUI).
- Fixed progress bar for parallel tempered sampling.
- Make sdist

------
This will be a bug fix release over 0.1.13. The Linux wheels weren't tested
enough and the default reflectivity calculation backend didn't work correctly
in 0.1.13.
- fix default backend calculation.

------
- synthesising of datasets (Data1D.synthesise) can now be repeatable by
providing a seed.
- use Github Actions to test and build macOS wheels across all the Pythons.
- The _creflect.abeles reflectivity calculation can release the GIL for a
large part of its calculation now. This will enable parallelisation using
either Processes or Threads. Processes still have the edge on speed at the
moment.
- change reduction code for ReflectNexus.phase_angle to use degrees, not
radians. (previously angles were returned in a mix of degrees/radians)
- Speed of reflectivity calculation is improved between 10 and 20 %. This is
achieved by the use of C99 complex arithmetic instead of C++ std::complex.
This improvement does not apply to Windows because it doesn't have a
C99 standard conformant compiler.
- Document how to save a model.
- macOS wheels (and CI testing) have the cyreflect openmp option activated.
- An openCL reflectivity calculation backend is added that can use a GPU.
- Added a `reflect.use_reflect_backend` function to choose between the backends
used to calculate reflectivity.

------
- allow master and slave chopper parameters to be ignored in PLP reduction.
- tunable t_0 offset in PLP/SPZ reduction.
- possibly_create_parameter accepts default bounds/vary/constraint. This fixes
a bug in MixedSlab.
- Don't require numpy be installed before setup.py can run.
- Speed up reflectivity calculation if there is solvation.
- reflect.choose_dq_type for finding out fastest mode of resolution smearing.
- User can now select resolution smearing approach in ReflectModel. Choosing
between 'pointwise' or 'constant'.
- Add progress bar for Curvefitter.fit() (requires tqdm being installed).
- A few micro-optimisations.
- Optimized pickling/unpickling of Bounds instances. This can lead to a huge
performance increase (~ 40%) when doing parallel sampling.
- MCMC sampling initialisation made reproducible