Add SVGD

[SamDuffield]

We should definitely add Stein variational gradient descent (paper, code)

SVGD requires a kernel specification. IMO we don't need to supply a suite of kernels ourselves (aside from maybe a default Gaussian kernel).

I think we should enforce a kernel signature like

eval = kernel(params1, params2, aux1, aux2, batch, **kernel_params)

where kernel_params are any kernel hyperparameters such as bandwidth.

To future-proof against more sophisticated kernels that e.g. could use info from the model call via aux1 and aux2.

Also we should think about how to support adaptive kernel_params updates like the median heuristic used in the SVGD [paper](https://proceedings.neurips.cc/paper/2016/file/b3ba8f1bee1238a2f37603d90b58898d-Paper.pdf.

[EtayLivne]

Hi Sam!

I recently got interested in this repo & Normal Computing in general. I wanted to contribute in some way so took a stab at this issue. I'm a bit uncertain about some of the math, so I thought I'd implement the 90% that I think I figured out, and ask about the remaining 10%.

Would you be interested in me opening a pull request with the code that I have? It has a unit test that shows it works in the sense that it adheres to the API, but I haven't tested the logic yet because of the parts I'm unsure of.

[SamDuffield]

@EtayLivne yes absolutely open a PR and we can discuss directly in the code. Thanks for taking a stab!!