This reduces a lot of memory copying and allocations, since
the clip stack built each frame now just contains handles
into the clip store, rather than copying the entire mask cache.