feat: add downstream task evaluation pipeline

[Sudhendra]

Summary

• add a downstream-task evaluation framework with offline benchmark prep for HotpotQA, QASPER, and DS-1000 plus deterministic scoring and baseline compressors
• add a runnable paired evaluation CLI for full-vs-compressed context comparisons, including explicit learned backend support for local MLX adapters and Tinker checkpoints
• document the downstream eval workflow and fix autoformat.sh so worktree pushes run hooks correctly

Test Plan

• pytest tests/test_downstream_dataset.py tests/test_prepare_downstream_eval.py tests/test_downstream_benchmarks.py tests/test_downstream_baselines.py tests/test_downstream_scoring.py tests/test_evaluate_downstream.py tests/test_evaluate_downstream_cli.py -v
• ruff check src/evaluation/downstream scripts/evaluate_downstream.py scripts/prepare_downstream_eval.py tests/test_downstream_dataset.py tests/test_prepare_downstream_eval.py tests/test_downstream_benchmarks.py tests/test_downstream_baselines.py tests/test_downstream_scoring.py tests/test_evaluate_downstream.py tests/test_evaluate_downstream_cli.py
• smoke-tested tiny downstream runs in the worktree and reviewed summary artifacts for baseline vs learned compressors

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 66d1107d4f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".