add system prompt and other cli args

[hjc-puro]

• We add cli args for test model system prompt, base url, api key, max tokens, and rpm rate limit
• We delay <think></think> stripping so that the full generation with reasoning is logged but the judge still sees only the answer post-

example command:

python eqbench3.py \
  --test-model "model-name" \
  --judge-model "claude-3-7-sonnet-latest" \
  --test-base-url "http://localhost:30000/v1" \
  --test-api-key "x" \
  --test-max-tokens 32768 \
  --iterations 3 \
  --runs-file "${RESULTS_DIR}/run.json" \
  --elo-results-file "${RESULTS_DIR}/elo.json" \
  --threads 4