[eb8680]

Addresses #821

This PR migrates handlers.llm from litellm to any-llm, as suggested in #821, removing litellm entirely as a dependency. For the most part it's a seamless migration because any-llm is designed for exactly this.

There is one annoying but important edge case which litellm handled but any-llm does not: OpenAI's first-party Completion API (via any_llm.completion) does not support reasoning when there are tools in the mix; OpenAI want you to use their Responses API instead (via any_llm.responses). This breaks uniform model access through any_llm.completion for OpenAI models.

The path of least resistance is to do what litellm.completion already does, which is recognize those models and translate the completion call to an equivalent responses call. The any-llm developers have said they have no plans to do this themselves, so it has to live in our code as a bunch of private helper functions in handlers.llm.harness.provision.any_llm.

Apart from this, the diff is mostly just boilerplate accounting for small API and behavior changes from the migration.

I think this PR is almost complete but it is a large breaking change and mostly LLM-authored so I want to run all the examples end-to-end and examine the outputs myself to make sure there are no subtle regressions I've missed, e.g. in the integration with langfuse.

[eb8680]

This can be reviewed and merged. I inspected some examples and everything seems to check out.

[eb8680]

I will rebase this on master once #824 lands