Fix #359 - Improve performance for many islands of constraints.

[cameronwhite]

For a modified ragdoll demo (35 rows of ragdolls and more spacing between them), this improved performance by about 30% overall. Inside of processIslands in the profiler, the time spent in solveGroup went from around 35% to >95%.

[erwincoumans]

Thanks for this effort, and excuses for the huge delay. I am cleaning up old PR's. Tried this on the Ragdolls Benchmark in the ExampleBrowser, disabling gravity, press 'g' (disable GUI) and hold 'p' for a while: recording detailed performance trace json file. Open the timings_0.json trace in Google Chrome using "about://tracing"). Comparison shows no benefit. Let't close it for now, unless you can show more convincing benefits. See old and new images, both around 3.7-3.9 ms.

OLD, around 3.7ms:

NEW (this patch applied) around 3.9ms:

[cameronwhite]

Hi Erwin,

Did you also adjust the ragdoll demo to create more constraints and islands? In BenchmarkDemo::createTest3(), I increased size to 35, and sizeX and sizeY to 2.0, and saw about 20ms spent in processIslands() with the original code, compared to 7.5ms with these changes.

[erwincoumans]

No I didn't. Let me try it again soon, and I'll report the timings. Thanks for getting back to me. Also, I will look at the other pull requests soon, there are too many outstanding.

[cameronwhite]

I forgot this PR was still open - I've rebased it on the latest upstream master and applied clang format

With the changes to the ragdoll example described in https://github.com/bulletphysics/bullet3/pull/386#issuecomment-329633730, I see a drop of about 10ms per timestep overall