* Switch all filter effects to run through the single source blend shader, instead of composite shader.
* Each render target uses a single alpha batcher now that the batching algorithm is reasonably efficient.
* Change tile size to 256x256. This seems to be a better tradeoff for GPU vs CPU time on all sites I tried.
* Switch blend targets to only allocate the used rect, rather than the entire tile size rect.

Closes #423.