Sandbox backends

[K-Jo]

Mainly Tensorflow (?) but other backends might need investigation as well, restricting I/O access.

[lantiga]

@K-Jo for GA?

[K-Jo]

@lantiga can we do a preliminary investigation on how much work this is before we decide? We can GA with recommendations.

[lantiga]

The bottomline is that we need to exclude io_ops (see https://www.tensorflow.org/api_docs/cc/group/io-ops) upon importing in MODELSET, or (maybe simpler for now) return an error from MODELSET if the graph contains any io_op.

ONNXRuntime and PyTorch (TorchScript included) do not support IO in the graph, they only deal with transformations of tensors, so there's no need to sandbox.