When flattening the display list, we currently pay a large memcpy overhead cost inside create_primitive. This is mostly due to the large enum struct sizes that are copied around on the stack. In some talos tests, this is ~35% of the time spent inside WR. We should be able to reduce this significantly by refactoring how create_primitive and primitive interning work in general, by using interners specific to the key / type, rather than a generic interner and primitive container.

This has landed in https://github.com/servo/webrender/pull/3336. The associated try run drops the dl_mutate talos performance to ~4700 (see https://treeherder.mozilla.org/#/jobs?repo=try&revision=7160b048de4bc64e76bee67a8b8401b5ec46217d&selectedJob=213316786).