(1/n) Refactor benchmark execution behind a shared runner interface

[kargibora]

This PR is one of the large refactors that is discussed, for implementing more proper task definition in which task is defined as a .yaml files that calls certain functions with set of values (like dataset revision, dataset...).

Summary

The current generate-and-evaluate flow is implemented directly inside generate_and_evaluate.py. Besides running
pairwise evaluation, this module is also responsible for selecting the workflow, resolving task baselines, constructing
generation settings, and initializing the judge.

This works for the current benchmarks, but adding a benchmark with a different execution flow would require adding more
task-specific conditions to the same module. The goal of this PR is to introduce a small runner abstraction before
adding further benchmark integrations.

What changed

This PR adds a BenchmarkAdapter contract that connects supported tasks to a benchmark runner. The existing pairwise
evaluation remains the only registered implementation and acts as the fallback, so the runtime behavior of existing
tasks is preserved.

The following responsibilities were also extracted from generate_and_evaluate.py:

• Native baseline definitions and lookup now live in judgearena/baselines.py.
• Shared generation argument construction lives in judgearena/benchmark.py.
• Judge model construction is shared between benchmark runners.
• ELO and pairwise evaluation now use the same generation configuration logic.
• The existing pairwise workflow is exposed as run_pairwise() and selected through the adapter registry.

The execution path is now:

CLI configuration
    -> benchmark adapter resolution
    -> selected benchmark runner
    -> generation and evaluation