ImageDecodeBench has a few places where it allows noise to dominate measurements:
* It runs all iterations and divides by iteration count to get the average. This takes long enough that the process is swapped out by the OS. This adds noise.
* It includes setup and teardown in the time, 
* It runs through all frames of an animated image and again averages. If an animated image must decode a large first frame but all subsequent frames are small sub-region updates, the average will be faster than the slowest decode and slower than the faster decodes. It will not represent what actually happened. For example, if decode times are improved that will be buried under the small frame times that don't spend much time in decode.

To more accurately represent measurements, we should time only decode time for each frame and report it separately. This allows us to trim out noise and make better comparisons.

Also, we should add support for timing only raster.