Assuming the axis has an `edges` attribute, I assume it should also have a
`widths` attribute, so I thought we can just use that for the pull bar widths
instead of calculating our own assuming a regular spacing (which might not be
true).

For the patch-width, we just use `right_edge - left_edge`, instead of what was
done before, which dividing this difference by the number of pulls and then
multiplying them again with the same number.

Necessary when doing pull plots for histograms with an axis of type
`hist.axis.Variable` with varying bin widths.

Co-authored-by: Henry Schreiner <HenrySchreinerIII@gmail.com>

Co-authored-by: Henry Schreiner <HenrySchreinerIII@gmail.com>

…egular-pull-widths

I add 2 tests

- `test_image_plot_ratio_variable_axis_with_regular_bins`:
  Same as `test_image_plot_ratio_callable`, but use an `axis.Variable`, but set
  the bin edges to be regular, same as former test with `axis.Regular`. With
  this I just check that the plot doesn't depend on the axis type. For this we
  re-use the existing baseline plot.

- `test_image_plot_ratio_callable_bayesian_blocks_bins`:
  Actually try the ratio plot with variable bin widths. For this I ran the
  bayesian blocks algorithm on 5000 random numbers. I used that instead of 1000
  numbers because for 1000 the binning seemed very coarse for me and the fit
  plots just looked like a triangle. For this we create a new baseline plot.

Also I added a helper function to create these `plot_ratio_callable` test for
different axes/binnings in order to avoid code duplication.

TODO: The baseline plots for
`test_image_plot_ratio_callable_bayesian_blocks_bins` looks like something is
off, the centre pull is much larger than the pull errorbars. I have to check
that the bin widths are correctly taken into account in the errorbars and ratios.

Most tests use `np.random.seed(42)` to re-seed the global RNG. Initially I didn't
follow this in my tests, because according to its documentation it is not
best-practice and a legacy function.

> The best practice is to not reseed a BitGenerator, rather to recreate a new
> one. This method is here for legacy reasons.

Instead, I created a custom RNG with a seed and used that to generate my random
numbers. However, the tests failed and I noticed that the random seed not only affects the numbers to
fill the histogram with, but also the fit, which I can't affect via a custom RNG,
so I have to re-seed the global RNG.

Therefore, I fall back to using `np.random.normal`.

Also update baseline with fix