It seems like the sizes for these data structures can be controlled from
Web content, and we are already prepared to deal with OOM conditions,
except that we are using infallible allocations by mistake.
This fixes Valgrind errors while running TestWebGLElementArrayCache.
This also fixes the way we do content/canvas/test in the Makefiles, which had been broken by my attempt to enable compiled-code tests there (bug 732660).
Since I still don't know how to do compiled-code tests there, I've disabled them. At least we have the rest of content/canvas/test back.
Let's hope that no regression happened.