Unify ArrowTensorType tables and Tensor blocks

[ericl]

Why are these changes needed?

This PR removes the special-case Tensor block format from Datasets, in favor of representing these as single-column Arrow tables.

Before:

ray.data.range_tensor(10000)
# -> Dataset(num_blocks=200, num_rows=10000,
#            schema=<Tensor: shape=(None, 3, 5), dtype=int64>)

After:

ray.data.range_tensor(10000)
# -> Dataset(num_blocks=200, num_rows=10000,
#            schema={value: <ArrowTensorType: shape=(3, 5), dtype=int64>})

Read/write numpy operations have been changed to support Arrow tables with tensor columns.

This both removes a lot of special-case Tensor handling and cleans up the user story for tensor support. In the future, we can add more methods to make working with Tensors simpler on top of this common path.

[scv119]

looks like a great simplification! I'd defer this to @clarkzinzow to review and accept.

[clarkzinzow]

This looks great! I'm glad that it turned out to be a pretty simple port.

[clarkzinzow]

Hmm should be easy to add, we can look at supporting this.

[ericl]

Ok, do you have any pointers? I can give it a try