Expand all saveable operations to generate a single C++ restore call.

This allows us to avoid repeated index lookups and perform a sequential scan of the index in the common case where we are doing a full restore, or a restore from a sub-model.  It also dramatically reduces excessive restore parallelism.

Testing with a checkpoint with 1000 100x100 tensors, restoring from CNS drops from ~1m to ~5 seconds.

PiperOrigin-RevId: 180827583