evaluator: preserve signed GSM8K ground-truth answers

[huangyz0918]

Summary

Preserve optional +/- signs when extracting GSM8K ground-truth answers, both after #### and in the plain-number fallback.

Bug

During a live OpenRouter smoke test, Gemini 2.5 Flash Lite correctly returned -3 for a temperature changing from 2 to -3. With reference #### -3, ProEval stripped the minus sign, compared against 3, and returned error score 1.0.

When a signed final answer followed earlier numbers in the solution, the #### pattern failed entirely and the fallback could select an unrelated earlier number.

The fix changes only the two extraction regexes. Existing comma handling and final-answer precedence are preserved.

Verification

• Real OpenRouter/Gemini calls after the fix returned -3 and -1234.5; both now receive error score 0.0, including a comma-formatted reference with earlier reasoning numbers.
• Existing repository suite: 123 passed.
• Local regression checks cover 31 extraction/precedence/scoring subcases.
• Only the two-line production fix is included. Test scripts, logs, API credentials, and result artifacts remain outside the PR.