This way is both kinder on event-based enumeration and less work to
unwind allocations for when properly implementing GLFW_OUT_OF_MEMORY.