Use the getRawSkeleton() function for better performance

[mxgrey]

On dart's speed_test example application, this improved performance by ~12% on my machine.

[azeey]

I tested this with one of the SubT practice tunnel circuits. Unfortunately I didn't see a noticeable performance improvement. This is probably because the time spent in step() is dominated by collision checking for this particular world.

[azeey]

Seems like there was some improvemen. Here is a plot from Remotery showing time spent in World::step.

I added category bits and ODE collision space hierarchies here, which seem to significantly improve
the performance of dSpaceCollide.