Implement SHM serialization for Tensors

[Vivek Panyam]

Maybe also make Handle serialize differently based on the size of the tensor? For large tensors, it uses shared memory. For small ones it serializes them inline? Make Handle a variant so we can choose dynamically.