Implement vectorize for XTensor Ops

[ricardoV94]

Some Ops are undefined, but most things are pretty straightforward.

The biggest challenge is when vectorization would imply new dimensions (say xtensor_from_tensor). There is no trivial way to allow the user to tell us what those dimension labels should be.

And sized operations (where numerical inputs determine output sizes). However those are equally tricky in tensor graphs.

I decided to implement most methods as an Op property. I think this should be the default, as it keeps code organized. Users can still implement their custom dispatch if it's missing, and we can offer a two level dispatch (not to accidentally override the error checking that is now done for all Ops before dispatching)

Missing tests

[ricardoV94]

I removed the public vectorize_node, because there's no point to it. If you want to vectorize a single node, just call vectorize_graph on the outputs. It was actually masking issues (one test is failing because of that), and forcing us to test cases that can't be generated from vectorize_graph at all.