This removes the output & mask caching done by graph networks in graph mode.

The primary function of the caching was to avoid recomputing the forward pass graph of the Model when querying its output mask. As a result, we change the way output masks are obtained in order to avoid redundant computations, by setting a tensor's mask as metadata on the tensor object (in graph mode) and relying on that metadata where applicable.

PiperOrigin-RevId: 207580457