Refactor/consolidate eval pipelines

[kargibora]

Consolidate evaluation pipelines

Summary

Both evaluate.py (evaluate_completions) and generate_and_evaluate.py (main)
independently build annotation DataFrames, compute preference summaries, serialize
results to JSON/CSV, and write run metadata — with slightly different code each time.
This PR extracts that shared logic into reusable helpers so both pipelines go through
the same path.

Depends on: #31

What changed

New / changed	Description
EvaluationResult dataclass	Pure-data container holding annotations_df, prefs, summary, and run_config. No I/O.
build_annotation_dataframe()	Builds a single DataFrame from forward (and optionally reversed) annotations, replacing duplicated concat-and-label logic in both pipelines.
build_evaluation_result()	Computes preference summary and packages everything into an EvaluationResult. Pure logic — no disk access.
save_evaluation_result()	Writes an EvaluationResult to disk (CSV + JSON + run metadata). Single place for all serialization.

Files touched (relative to PR 1)

• judgearena/evaluate.py — Added the four helpers above; refactored evaluate_completions to use them (+220 / −6)
• judgearena/generate_and_evaluate.py — Replaced ~60 lines of inline DataFrame building, summary computation, JSON writing, and metadata writing with calls to the new helpers (+39 / −55)

Before → After

Before: each pipeline had its own copy of:

df = pd.DataFrame(annotations)
df["instruction_index"] = ...
df["model_A"] = ...
# ... swap_mode == "both" handling ...
df.to_csv(...)
summary = compute_pref_summary(prefs)
json.dump(...)
try:
    write_run_metadata(...)
except OSError: ...

**After:**
```python
annotations_df = build_annotation_dataframe(...)
eval_result    = build_evaluation_result(annotations_df=..., prefs=..., run_config=...)
save_evaluation_result(eval_result, output_dir=..., ...)

[ErlisLushtaku]

nit: this explanation might be redundant