fix: improve the lexico-minimum finder in LemkeAlgorithm.

[ShoYamanishi]

The following issues in btLemkeAlgorithm::findLexicographicMinimum() are addressed in this commit.

Logic Issues

• The z_0 column located at A(,2dim), is used for lexico-minimum test, which should not be used.
• If the check on the 1st column 'q' produces multiple minimums due to degeneracy,
  and if it contains z_0, the tie must be broken in favor of z_0 and return,
  which is not performed in the original.

Efficiency Issues

• The 1st part to construct 'Rows' is unnecessary.
• The 2nd part has three nested loops with brute-force search, which is inefficient.

Description of the Fix
The improved version performs the lexico-minimum test column-by-column
from the left-most column of (-B^{-1}q:B^{-1}).
If the test finds a single row, it returns it.
If the test finds multiple rows and if they contain z_0, it returns z_0.
Otherwise, it moves on to the next column.
Please see 2.2.2 "Pivot Step" of
K. Murty, "Linear Complementarity, Linear and Nonlinear Programming" Heldermann Verlag, Berlin, 1998
for the algorithm details.

Test Results
Some performance tests are done and non-negligible improvement was observed.
Please see https://github.com/ShoYamanishi/AppleNumericalComputing/tree/main/14_lcp for the test details.

[erwincoumans]

Thanks a lot for sharing the improvements and the test cases. Let's merge it. Also, your Lemke implementation seems to perform well! By default, we use an iterative constraint solver in Bullet, and for joints we typically use reduced coordinates (Featherstone articulated body algorithm), so the Lemke implementation hasn't received much attention.