EHN: Add Numba-jitted linprog solver

[oyamad]

Supercedes #413

[pep8speaks]

Hello @oyamad! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-04-30 05:45:46 UTC

[coveralls]

Coverage increased (+0.09%) to 94.393% when pulling ba02284 on oyamad:linprog_simplex into 83ae0a2 on QuantEcon:master.

[QBatista]

@oyamad I've updated the default PivOptions parameter values.

Summarizing the motivation behind this choice, the recommendation in Linear Programming Computation is the following:

fea_tol = 1e-6
1e-11 < tol_piv < 1e-7

Based on the experiments on the Netlib problems, the choice of tol_piv affects the performance of the solver significantly, while fea_tol and tol_ratio_diff only marginally affect performance. It seems that there's a negative relationship between performance and tol_piv on these problems. The best parameter combination that I found is fea_tol=1.e-06, tol_piv=1.e-04, tol_ratio_diff=1.e-13]).

As a compromise between these experiments and the book's recommendation, we can set the default values to fea_tol=1e-6, tol_piv=1e7, tol_ratio_diff=1e-13.

[QBatista]

This looks great to me.

[mmcky]

thanks @oyamad and @QBatista. I didn't see this issue until just now with the ready tag. @oyamad please feel free to merge this when you're ready. I can also issue a new release when you need one.