Comparative benchmarking results

[dmbates]

I noticed that some of the benchmarking results show dramatic decreases in execution time and memory usage from the reference to the test runs. I believe this is artificial and that the the reference and test runs are on the same version of Julia and MixedModels. The difference is that the reference is run first and has the compilation overhead.

[dmbates]

I should have said Target and Baseline instead of reference and test.

[palday]

Yeah, I've noticed that the comparisons are extremely noisy as well. I stalled in #711 trying to find a set of benchmarks that are less noisy. It may be worthwhile to really reduce the portions that are actually being benchmarked, e.g. a call to updateL! instead of the entire fitting process.

[dmbates]

Agreed that benchmarking updateL! for a couple of large, complex examples will be more meaningful.