Rewrite macros using staged functions

[mlubin]

Major revamp of expression parsing macros using staged functions (requires Julia 0.4).
Primary new feature is that this lets us handle commutivity (#99) without sacrificing performance, and also we'll get faster quadratic expression building.

Before:

$ julia speed.jl 
PMEDIAN BUILD MIN=0.311928322  MED=0.330600015
PMEDIAN WRITE MIN=1.731297123  MED=1.751828376
CONT5 BUILD   MIN=0.179188774  MED=0.216120392
CONT5 WRITE   MIN=1.45397678  MED=1.47033628

After:

$ julia0.4 speed.jl 
PMEDIAN BUILD MIN=0.329470738  MED=0.332986915
PMEDIAN WRITE MIN=5.201167811  MED=5.25195727
CONT5 BUILD   MIN=0.232650668  MED=0.237446502
CONT5 WRITE   MIN=3.984346782  MED=3.994990259

So 5% slowdown for PMEDIAN build time, and 30% slowdown in CONT5 build time. This is a reasonable trade off for the extra flexibility, I think. I can do a bit more tweaking to see if I can get extra speed ups, and we can expect improvements in compiler optimizations for staged functions. Slow down in write time is unrelated, should investigate separately what's causing this in Julia 0.4.

Generated code is a bit more type stable now, since new intermediate variables are created for the output of each subexpression.

TODO:

• Improved handling to speed up quadratic expressions (requires quadratic benchmarks)
• General speed tweaks
• Julia 0.3 compatibility
• Compatibility with extensions

[joehuchette]

:) by the quadratic handling stuff do you mean doing the expression tree search to detect quadratic vs. linear expressions?

[mlubin]

No, just adding a case in addToExpression_reorder that separates the two quadratic terms from the coefficients, right now I'm being lazy by multiplying one of the quadratic terms with all of the coefficients first.

[IainNZ]

Brave new world :) Feels so much more right than the old approach

[mlubin]

I still kinda like that you could read a model and know which terms are coefficients and which are variables, but that doesn't generalize very well to nonlinear expressions. Should we have some sort of a style guide?

[mlubin]

Rebased, compatible with 0.3 now. @IainNZ, does this break macros for JuMPeR?

[mlubin]

@IainNZ, just tested JuMPeR with this and tests pass. Safe to merge?

[IainNZ]

Yep, go for it.

[tkelman]

The Maros-Meszaros test set is probably the standard benchmark for QP's. If there isn't a Julia QPS reader yet (?), they're available as mat files here https://github.com/YimingYAN/QP-Test-Problems (I did the exact same thing a few years ago - used a mex interface to CoinUtils to read them into Matlab). Or maybe not too hard to parse from the originals http://www.cuter.rl.ac.uk/Problems/marmes.shtml

These are a subset of CUTEr of course. Wonder where the Julia CUTEr/CUTEst interface got to.