Microcanonical Langevin Monte Carlo

[reubenharry]

This PR adds the Microcanonical Langevin Monte Carlo to Blackjax, following the discussion of issue #2035.

A couple things to resolve:

• This incurs a Blackjax dependency. The only alternative would be to copy all the relevant Blackjax code into the implementation, which doesn't seem like a great idea.
• The code is in the main directory, because if I try anything in the infer directory, such as hmc.py, I get an error ImportError: cannot import name 'device_get' from partially initialized module 'jax' (most likely due to a circular import) (/opt/homebrew/Caskroom/miniconda/base/envs/mclmc/lib/python3.11/site-packages/jax/__init__.py). Does this look familiar?
• Before merging, I would probably like to set some better defaults for the amount of tuning etc, for robust performance.
• I should presumably add a simple test

[juanitorduz]

Suggestion: What about adding BlackJax as an optional dependency and adding this component intro de contrib submodule?

And it would be great to have this simple example visible in the example gallery 🙂

[juanitorduz]

hi @reubenharry , do you need some support with this one? I could also offer to branch out from this branch and keep working on a PR :)

[reubenharry]

Hi, sorry for the slow response. Yes, if you're interested in doing that, it would be extremely helpful - I wasn't sure how to proceed.

[juanitorduz]

Hi, sorry for the slow response. Yes, if you're interested in doing that, it would be extremely helpful - I wasn't sure how to proceed.

FYI #2124 (comment) :)