[Datasets] Make ragged tensor shape more descriptive

[bveeramani]

Description

>>> arrays = [np.zeros([256, 256, 3]), np.zeros([224, 224, 3])]
>>> df = pd.DataFrame({"image": TensorArray(arrays)})
>>> ray.data.from_pandas(df)
Dataset(num_blocks=1, num_rows=2, schema={image: TensorDtype(shape=None, dtype=float64)})

I expected the TensorDtype shape to be (None, None, 3) but I got None instead.

Use case

We need this functionality to make to_tf work with ragged tensors. If we make the ragged tensor shape more descriptive, we can accurately infer the output_signature for tf.data.Dataset.from_generator.