slice_numpy_mcmc tuning fix

[gmoss13]

What does this implement/fix? Explain your changes

fixes inf values being returned by the slice_numpy mcmc sampler

Does this close any currently open issues?

Fixes #1060. Also related to #910

Any other comments?

The issue was not in the sampler "diverging", c.f. #910:

The main reason seems to be the number of warmup steps. If it is too small, chains have not converged and samples diverge (why?)

Instead, we generated the samples to return initially with np.empty(), but did not account for the tuning of the bracket width, so we never defined the first entries of the sampels returned. This is why increasing warm up steps also fixed the issue - the nan values were completely removed if the warm_up steps were increased enough.

[gmoss13]

As a side note, the code for the slice sampler is not very well documented, so it was time consuming for me to find this bug because first I had to decipher the code. Could add some documentation here in this PR?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 72.91%. Comparing base (8efb8e4) to head (ec2c4da).

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1174       +/-   ##
===========================================
- Coverage   83.69%   72.91%   -10.79%     
===========================================
  Files          93       93               
  Lines        7397     7399        +2     
===========================================
- Hits         6191     5395      -796     
- Misses       1206     2004      +798     

Flag	Coverage Δ
unittests	72.91% <100.00%> (-10.79%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files	Coverage Δ
sbi/samplers/mcmc/slice_numpy.py	92.05% <100.00%> (+0.06%)	⬆️

... and 24 files with indirect coverage changes

[janfb]

Great catch! 👏
Absolutely, if you could add some documentation that would be awesome

[janfb]

Looks good! thanks for fixing this! 👏