Convergence testing based on the Rubin-Gelman statistic is implemented
via `setup_convergence_test`. This does not truncate the search yet, but
simply adds a trace to the walkers plots for testing

- Addresses incorrecelt named attributes and prevents error if
  sftfilepath is not specified

Also adds thetas_at_ref data to the Writer object

Now forces use of dtglitch = None for no glitches

1) Adds 'add_noise' option (defaults to true)
2) Better checking of Writer for if all glitch parameters have the same
length (avoid strange errors)
3) Fixes issue in which the tqdm ignore command was itself ignored

Transforms the single pyfstat.py into a python module splitting the
relevant code into separate sub-files in pyfstat. This should result in
improved readability.

1) Update README installation instructions
2) Put all set_up code into seperate named functions
3) Remove underhanded mpl backend setter

Prior to this, the analytic fits data was all for twoF = 60, this
updates this and adds a tag to the filename to make it it clear

Use N=1000 for the recoveries and N=10000 for the test in noise

Issues:
1) If tstart and duration where not explicity searched over, a loop
condition would quickly send the duration to infinity
2) Fixed half norm definition

1) Changes the way the prior is built to be wider than the initial box
2) Fixes error where the F0 and F1 where always chosen in the middle of
the box