Implement a per-thread color space conversion cache

This makes the cross-colorspace painting about 2 times faster.
There is really high contention on this mutex, so the per-thread storage
relaxes it a bit.