NMObjectCache was assuming there would never be more than one object
with the same path, but since NMClient is an NMObject, it was getting
cached too, so if you created two clients and then unreffed one of
them, it's possible the wrong one could get left in the cache, causing
a crash the next time the other one called nm_object_cache_clear().

Fix this by only adding NMObjects to the cache in the codepaths where
we also check to see if the object was already in the cache.

(This also means we can remove the "except" argument to
nm_object_cache_clear(), since the NMClient won't be cached any more.)