Instead, bump the pixmap's refcount at the bottom of post_damage to reflect the compositor's hold on the buffer, and "destroy" the pixmap in the buffer release callback (which will dec the pixmap's refcount and free if necessary).
cc @ofourdan
Inspired by !520 (closed), but unlike !520 (closed) this mostly fixes the memory leak for me. I'm not completely convinced it's leak-free in the SetWindowPixmap path, but at least I'm no longer leaking one stream per resize event.