Change default tolerances to be a bit more realistic

[denehoffman]

The default epsilon values for basically all the methods were a bit too precise. While they work fine for the Rosenbrock demos, most real-world applications don't start with that level of precision (scipy uses 1e-4 for both tolerances on their Nelder-Mead and esp * 1e7 for their 64-bit float tolerance for L-BFGS-B, for example). I've tried to scale these and use values that can work with either 64- or 32-bit floats and approximate those used by the 64-bit only existing libraries (scipy and the FORTRAN implementation of L-BFGS-B). As a result, tests needed to be updated (in precision only) and benchmarks will probably run faster simply because it will take fewer steps to converge.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 70.70%. Comparing base (63b6e11) to head (f9bc385).
Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #44      +/-   ##
==========================================
+ Coverage   70.37%   70.70%   +0.32%     
==========================================
  Files           8        8              
  Lines        1826     1826              
==========================================
+ Hits         1285     1291       +6     
+ Misses        541      535       -6     

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

[denehoffman]

This seems to improve coverage too, but only because assert!(approx_eq!(...)) doesn't seem to provide the same coverage as assert_approx_eq!(...)! I guess the LLVM coverage thinks there's a path where the inside fails but the outside doesn't? Strange.

[github-actions]

Benchmark for 7b83eef

Click to view benchmark

Test	Base	PR	%
nelder-mead: rosenbrock/Adaptive/2	32.6±0.77µs	21.4±0.74µs	-34.36%
nelder-mead: rosenbrock/Adaptive/3	83.6±1.17µs	56.9±1.13µs	-31.94%
nelder-mead: rosenbrock/Adaptive/4	165.5±1.63µs	37.7±0.71µs	-77.22%
nelder-mead: rosenbrock/Adaptive/5	267.6±5.52µs	186.9±4.27µs	-30.16%
nelder-mead: rosenbrock/Standard/2	32.6±0.31µs	21.3±0.24µs	-34.66%
nelder-mead: rosenbrock/Standard/3	69.6±1.01µs	47.4±0.58µs	-31.90%
nelder-mead: rosenbrock/Standard/4	144.9±3.95µs	32.1±0.50µs	-77.85%
nelder-mead: rosenbrock/Standard/5	238.4±4.30µs	43.5±0.93µs	-81.75%