fix(autofix): Cache repo client objects to avoid re-initialization

[roaga]

See this profile.

We were creating new repo clients over and over again, and each time, it added 1-1.5s.

• before invoking a step, ~1.5s redundant
• pre-filling coding step memory: ~5s (more for more files)
• every expand_document tool call: ~1s

This PR stores created repo clients in memory in the context so that throughout a single step, we only need to initialize one repo client for each repo.

[trillville]

🚀

[roaga]

Callout from Ram: is the RepoClient holding open any connections? If so, that would be a problem.

However, I think it's not doing that. It's just wrapping HTTP requests.

But @jennmueng can you confirm?

[jennmueng]

Nice find! functools.cache might be a simpler solution. Also, yeah it should just be http requests from the github client!