Optimize clipping

Currently, we're using just a simple alpha mask, unlike Skia, which has a very complicated clipping algorithm. It's like 5000 LOC. But this algorithms is always faster that just a mask and uses less memory.

Begin by comparing the current alpha-mask clipping approach with Skia's clipping algorithm, then measure their speed and memory use. The issue names no files or tests, so the implementation entry point and validation coverage must be located first. Done means clipping is faster and uses less memory without changing rendered output.

Written by the indexing model from the issue text.