Add pre-sorting of levels to RFP

[palday]

Turned out to be less difficult than I originally thought.

[dmbates]

Don't you mean that nnz(spL) counts the non-zeros?

[palday]

oops, yes.

ee1bf47

[dmbates]

I am getting a test failure of

Test Summary:        | Pass  Total  Time
likelihoodratio test |   22     22  5.4s

Poisson: Test Failed at /Users/dmbates/.julia/dev/MixedModels/test/bootstrap.jl:65
  Expression: isapprox(gm4.β, gm4sim.β; atol = norm(stderror(gm4)))
   Evaluated: isapprox([0.5293132302958539, 1.1072973757597735, -0.91841266986093, -0.022191955068345592], [0.4584581329384717, 1.2388751637383038, -0.5251807690359246, -0.02264331510499933]; atol = 0.3972560419906301)

[palday]

I am getting a test failure of

Test Summary:        | Pass  Total  Time
likelihoodratio test |   22     22  5.4s

Poisson: Test Failed at /Users/dmbates/.julia/dev/MixedModels/test/bootstrap.jl:65
  Expression: isapprox(gm4.β, gm4sim.β; atol = norm(stderror(gm4)))
   Evaluated: isapprox([0.5293132302958539, 1.1072973757597735, -0.91841266986093, -0.022191955068345592], [0.4584581329384717, 1.2388751637383038, -0.5251807690359246, -0.02264331510499933]; atol = 0.3972560419906301)

Interesting. That's not the best test, but I've loosened the tolerances and kept it in for now (2c230e6). That test is supposed to test the simulation code and show that we get similar-ish parameter estimates when we fit a model on simulated data. I've set the tolerance to 2 * norm(stderror(gm4)), which I've hand-waved as vaguely corresponding to a 95% confidence region.