add head_array() methods for unstructured computation of heads

[dbrakenhoff]

• steady and transient models
• make headgrid use head_array under the hood.

[dbrakenhoff]

test fail fixed in #138

[dbrakenhoff]

@mbakker7 this is ready for review now

[mbakker7]

This looks nice. I tried the headgrid function in timflow.transient on a Macbook Air. It only has 10 cores (4 for performance, 6 for efficiency). With parallel=True I get a speedup of 10%. With parallel=4 I get a speedup of 30%. Maybe I don't have enough elements to worry about, as the run time is still short. Computing a 50 by 50 grid for 100 times only takes about 3 seconds. Is a bigger improvement expected when the model is so large that it would take, say 30 seconds?

[mbakker7]

Just ran a more substantial example (the parallel_example from #140). Generating a grid is 3 times faster with parallel=True, so great job! Good note that there is hardly any time lost by computing the grid at 10 different times rather than just 1 time.
So nice job!

[dbrakenhoff]

Yea, i expect the speedup to be more significant for larger models (more elements) and finer grids, up to a certain point of course.

I think 3x speedup on 4 main cores seems reasonable. On my PC using the maximum number of threads (mp.cpu_count()) also didn't get me the fastest run time. The optimal setting probably depends a bit on the problem and the overhead of parallelization.