Preserve evaluator feedback for undecodable score files

[Sylvester Kaczmarek (sylvesterkaczmarek)]

Fixes #1457.

Read the diagnostic preview of scores.csv as UTF-8 with replacement for undecodable bytes in the workflow, pipeline, and runner evaluators. CSV parsing remains strict, and corrupt output still produces negative feedback instead of a second decoding exception.

Adds 15 parameterized cases covering invalid/truncated UTF-8, empty files, malformed Unicode CSV, and valid scores across all three evaluators. The six UTF-8 corruption cases fail against the original code. Tests also verify that existing feedback and file contents are preserved.

Validation on macOS arm64, Python 3.11:

• 26 selected offline tests passed, including all 15 new cases.
• New tests pass the pinned Ruff lint and formatting checks; production files introduce no additional Ruff diagnostics relative to main.
• Changed modules compile; git diff --check passes.

Model calls, execution environments, and submission checks are mocked. Full research/training workflows were not run.

---

📚 Documentation preview 📚: https://RDAgent--1483.org.readthedocs.build/en/1483/