Add OpToNestedTensor (Prototype, for ref)

[SagiPolaczek]

(NOT FOR MERGE)
Hey All,

Following @YoelShoshan's proposal and #219 I implemented some sort of prototype for supporting Nested Tensors.
Note that it requires latest PyTorch version (and it will to do so, since they still working on it). Current version is 1.13.0.

Example of how I used it in a data pipeline (DTI task, so the sequence are differ in size between samples!):

        dynamic_pipeline = [
            # read data
            (OpReadDataframe(data=df, rename_columns=rename_columns, key_column=None), dict()),

            # dummy encoding (ASCII values) and conversion to tensors
            (OpDummyStringToNumbers(), dict(key_in="data.drug.smiles", key_out="data.drug.encoding")),
            (OpDummyStringToNumbers(), dict(key_in="data.target.sequence", key_out="data.target.encoding")),
            (OpToNumpy(), dict(key=["data.drug.encoding", "data.target.encoding"], dtype=np.float32)),
            (OpToTensor(), dict(key=["data.drug.encoding", "data.target.encoding"], dtype=torch.float)),

            # covert to nested tensor
            (OpToNestedTensor(), dict(keys_in=["data.drug.encoding", "data.target.encoding"], key_out="data.input.nested_tensor")),
        ]

[SagiPolaczek]

for future ref:
The use of Nested Tensor should be consider to be in the collate phase, and not as a stand-alone op.

thanks @mosheraboh

[SagiPolaczek]

Cleaning PRs