It's very hard to write a proper micro benchmark. I wouldn't trust one written in AS3, unless AS3 was my target.
Here's a quick benchmark in Java. The LUT is purposefully padded so each entry is in its own big array, making this the worst possible scenario for the LUT (no cache):
http://pastebin.com/TnJJs4mQ
MathUtils.sin 321.727177
sinQuadratic 296.255125
Here is the same benchmark, but without padding the LUT. This is the best possible case for the LUT (fits in cache):
http://pastebin.com/j8MHiLzG
MathUtils.sin 27.967476
sinQuadratic 296.162242
The LUT has an array access and the low accuracy quadratic version has branching. They look pretty similar when not making use of the CPU cache and the LUT is most definitely faster when the cache can be used. From these results it looks like the LUT is better, since it has either the same or much better performance.
A real benchmark would need to use actual skeletons, since that is where we are using sin/cos. Doing that, I'm pretty sure we'd find that so much other work is being done, the speed of cos/sin is just not very important because it is not the bottleneck. Skeletons have relatively few bones. A complex skeleton is maybe 75-100 bones, so even with many on-screen at once, it's a pretty small number of sin/cos calls. Most of the work happens with the rest of the bone transform calculations and when applying timelines.
