Fixing Nelder-Mead in box optimization

[lalondeo]

This fixes the issue raised in #912. When Nelder-Mead's initial state is created, some of the points in the simplex lie outside of the box, so that the value of the corresponding logarithm is infinite. Since the barrier parameter is initially zero, this results in the objective's value at those points being zero * Inf = NaN, which trips up the optimizer. Starting out with a small but nonzero barrier parameter instead fixes the issue (resulting in the objective being infinite outside the box) and the code snippets that were given in the aforementioned post now give the correct results. However, this is more of a band-aid than anything else, as it seems reasonable to think that the points in Nelder-Mead's simplex should always satisfy the box constraints. Enforcing this would require changing large portions of the code.

[pkofod]

Thank you, I will have a look.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 84.80%. Comparing base (78ab1f4) to head (de4a317).
Report is 2 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1080      +/-   ##
==========================================
- Coverage   84.85%   84.80%   -0.06%     
==========================================
  Files          46       46              
  Lines        3480     3481       +1     
==========================================
- Hits         2953     2952       -1     
- Misses        527      529       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.