Add two variants of the KCI test

[OliverSchacht]

This PR is updating #201 so it should be easier to merge.

Original Text:
I am adding two variants of the kernel-based conditional independence test, that are aiming at improved computational efficiency.

Namely:

• The Randomized Conditional Independence Test (RCIT) translated from this R code.
• The Fast and Parallel Kernel-based Conditional Independence test (FastKCI), implemented based on an idea of @Biwei-Huang and myself.

I also added two easy unit tests checking the implementation.
I added the algorithms to the cit class, so they can be used with the PC algorithm for causal discovery.

Happy to hear your feedback!

[MarkDana]

Hi @OliverSchacht thanks for your amazing contribution! I will review this PR within next week.

[MarkDana]

Dear @OliverSchacht I really appreciate this contribution!

I reviewed the codes and the logic seems good to me. I didn't identify major running issues. The two unittests passed from my end, though I didn't do other accuracy benchmarks (e.g, those in TestPC.py/test_pc_simulate_linear_nongaussian_with_kci).

(One minor point is that I see KCI.py is changed on null_samples -> self.null_samples. This change seems unrelated. Though trivial, I revert it just for sanity check. Two unittests still passed after this reversion. But please correct me if i'm wrong.)

For now I think this pr is ready to be merged. One suggestion: could you please also add documents to https://github.com/py-why/causal-learn/tree/main/docs/source/independence_tests_index, with general instructions on using FastKCI and RCIT, with parameter choices on e.g., K, J. This could be done maybe after the working paper is finished.

Thank you again :)

[kunwuz]

Hi @OliverSchacht , thanks so much for your contribution! When you think it is ready, could you please write a short document as @MarkDana suggested? This would be very helpful for our users to start using this excellent test.