Cleanup and performance improvements

[rdeits]

Better version of #19

This does the following:

1. Replaces the non-const globals user_f and user_j with const Refs, using FunctionWrappers to ensure type-stability
2. Removes all the custom sparse_matrix code, because Julia's built-in SparseMatrixCSC does the same thing
3. Removes additional solveMCP methods which exist only to provide default arguments, because default argument values already create those extra methods automatically.

I did some profiling with the following benchmark:

using PATHSolver
using BenchmarkTools

function construct_problem()
    srand(1)
    n = 50
    M = sprandn(n, n, 0.1)
    M += M'
    q = randn(n)
    myfunc(x) = M*x + q

    lb = zeros(n)
    ub = fill(100.0, n)
    myfunc, lb, ub
end

myfunc, lb, ub = construct_problem()
results = @benchmark solveMCP($myfunc, $lb, $ub) setup=(options(convergence_tolerance=1e-2, output=:no)) evals=1
display(results)

Before this PR, I saw about 2.67 ms, and after this PR I see 2.18 ms, or about a 20% improvement in total time. That time includes all the time spent in PATH and computing jacobians, so the actual improvement in callback overhead should be better than 20%.

[coveralls]

Coverage decreased (-4.3%) to 95.652% when pulling 51a9e2b on rdeits:cleanup into 7e4928a on chkwon:master.

[rdeits]

Ok, I've also fixed compatibility with Julia v0.5, v0.6, and nightly on Linux. Appveyor is broken with Julia nightly due to JuliaLang/julia#20899

[coveralls]

Coverage decreased (-3.8%) to 96.25% when pulling 5c49e40 on rdeits:cleanup into 7e4928a on chkwon:master.

[chkwon]

This looks cool! Thanks for PR.

[coveralls]

Coverage decreased (-4.2%) to 95.833% when pulling eb13c31 on rdeits:cleanup into 7e4928a on chkwon:master.

[coveralls]

Coverage decreased (-4.3%) to 95.652% when pulling 700436c on rdeits:cleanup into 7e4928a on chkwon:master.