fix calculate_fd_gradient to deal with design pixel weights near 0 or 1

[mawc2019]

The function calculate_fd_gradient does not return the correct finite-difference gradients when the design pixel weight is 0, 1, or too close to them. This PR replaces the central difference with forward or backward differences in these circumstances, and also forbids a too large (≥0.2) step size for finite difference.

[codecov-commenter]

Codecov Report

Merging #2073 (5d22b7e) into master (5275f43) will decrease coverage by 0.05%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master    #2073      +/-   ##
==========================================
- Coverage   73.25%   73.19%   -0.06%     
==========================================
  Files          17       17              
  Lines        4902     4906       +4     
==========================================
  Hits         3591     3591              
- Misses       1311     1315       +4     

Impacted Files	Coverage Δ
python/adjoint/optimization_problem.py	55.32% <0.00%> (-1.15%)	⬇️

[stevengj]

Sounds good. How did you test this?

[mawc2019]

I tested it with uniform design patterns such as np.zeros(Nx*Ny), np.ones(Nx*Ny), 1.5*np.ones(Nx*Ny), etc. The adjoint and finite-difference gradients were compared.