Huge overhead using JuMP

[vepiteski]

When solving large LCP models using JuMP, I observe a huge overhead. Here is an example, the LCP is the Murty model.

`using PATHSolver
using LinearAlgebra
using Complementarity

MA=2.0*ones(n,n)
M=tril(MA,-1) + I

x=zeros(n)
q = vec(-ones(n,1))

mJuMP = Model(PATHSolver.Optimizer)
@variable(mJuMP, x[1:n] >= 0)
@constraint(mJuMP, M * x .+ q ⟂ x)

@time optimize!(mJuMP)
`

Its output trailer is as follows for n=1024

`Total Time. . . . . . . 0.018557
Residual. . . . . . . . 0.000000e+00
Postsolved residual: 0.0000e+00

5.496285 seconds (12.75 k allocations: 68.493 MiB)
`

As you may observe, @time returns 5.49 seconds and PATH itself needed 0.02 seconds. The overhead is much worse for larger models, for instance, for n=2048, @time returns 109.05 seconds but PATH itself reports 0.06 seconds. Actually, one has to wait that long before PATH begins its execution.

[odow]

Fix is in #122. I'll get a new release today.