Streamline run_inference_algorithm and the streaming average

[reubenharry]

The run_inference_algorithm code currently has a mode to incrementally compute a value (e.g. an incremental expectation), which saves memory. This PR presents a cleaner (and in my opinion, more functional) version of the code. The core idea is that to incrementally compute an expectation, the right thing to do is to modify the kernel, and then pass that kernel to run_inference_algorithm, rather than modify run_inference_algorithm. As part of this change, we also let transform be a general function on (state, info), which allows information from the info to be used (which we have needed).

• There is a high-level description of the changes;
• The branch is rebased on the latest main commit;
• pre-commit is installed and configured on your machine, and you ran it before opening the PR;
• There are tests covering the changes;
• The doc is up-to-date;