added the btNNCGConstraintSolver, based on the paper "Nonsmooth Nonlinear Conjugate Gradient Method for interactive

contact force problems". The solver needs a lot of iterations, before the quality goes up (~ 1000)
Thanks to Gabor PUHR for the contribution!
Improved the btLemkeSolver.
Remove the sparse optimizations from the btMatrixX.h, replace it with explicit call to rowComputeNonZeroElements (only used in the btSolveProjectedGaussSeidel), it was likely slowing things down, without being useful.
Re-enable SIMD in the solver (was accidently disabled in Bullet 2.82 release)

src/BulletDynamics/MLCPSolvers/btSolveProjectedGaussSeidel.h

@@ -20,11 +20,17 @@ subject to the following restrictions:
 
 #include "btMLCPSolverInterface.h"
 
+///This solver is mainly for debug/learning purposes: it is functionally equivalent to the btSequentialImpulseConstraintSolver solver, but much slower (it builds the full LCP matrix)
 class btSolveProjectedGaussSeidel : public btMLCPSolverInterface
 {
 public:
 	virtual bool solveMLCP(const btMatrixXu & A, const btVectorXu & b, btVectorXu& x, const btVectorXu & lo,const btVectorXu & hi,const btAlignedObjectArray<int>& limitDependency, int numIterations, bool useSparsity = true)
 	{
+		if (!A.rows())
+			return true;
+		//the A matrix is sparse, so compute the non-zero elements
+		A.rowComputeNonZeroElements();
+
 		//A is a m-n matrix, m rows, n columns
 		btAssert(A.rows() == b.rows());