feat: allow KVFlash paging on MLA K-only caches

Ling/bailingmoe3 stores compressed MLA K and reconstructs V, so the
pager no longer requires a V tensor on every attention layer.