Cache (4/5): Adopt inference caching in task runners

[ErlisLushtaku]

Description

Adopts the inference cache in the task runners through --run.store_root.

• Pairwise, meta-eval, MT-Bench, and Elo generation and judging share the do_inference cache.
• Elo temperature calibration uses the same judgement cache.
• Cache rows get instruction, model pair, and orientation metadata at the direct inference call.
• Dataset-provided completions stay as direct inputs when they do not run inference.
• Removes ignore_cache, cache_function_dataframe, and the old cache tokens.

This is stacked on #123.

[kargibora]

Overall this PR looks quite clear. It wires all the things that have been implemented and removes the legacy code. It is good that we got rid of many lines that were making the code harder to parse!

[kargibora]

Nice that we get rid of this!

[kargibora]

I also did not understand the reason of cache_metadata. Do we need it in , do_inference(), annotate_battles() and judge_and_parse_prefs()?

[ErlisLushtaku]

@kargibora Yes, it needs to reach all three in the current flow. It is not part of the cache key, but rather the info that doesn't necessarily affect determine the output e.g. instruction_id, model_a, orientation etc. do_inference uses it only when storing a miss, annotate_battles keeps it aligned with the rendered prompts, and judge_and_parse_prefs swaps model_a/model_b and orientation for the reversed prompt. I renamed it to cache_row_metadata to make this distinction clearer.

We could avoid passing a separate argument by creating an object such as:
InferenceInput(payload=prompt, cache_row_metadata=metadata)

What do you think?

[kargibora]

Thanks for making it clear! I think that can work