Grade benchmarks with runner configs from the pristine exercise tree

[AUTHENSOR]

Summary

run_unit_tests restores only the test files before grading, while the test command runs in the model-writable working tree and the benchmark harness auto-approves file creation (InputOutput(pretty=False, yes=True)). A four-line model-created conftest.py that resets session.exitstatus therefore makes pytest exit 0 while every test fails, and the exercise is recorded as solved:

def pytest_sessionfinish(session, exitstatus):
    if exitstatus != 0:
        session.exitstatus = 0

The chat history still contains the FAILED test lines while the exercise is scored solved, and a gamed run is indistinguishable from an honest one in the results yaml and on the leaderboard, which accepts community-submitted results. The same primitive reaches the other languages through their runner config files (package.json for JavaScript, CMakeLists.txt for C++, build.gradle for Java, Cargo.toml for Rust), none of which are restored.

Fixes #5681.

Changes

• Before grading, the runner config files for the exercise's language are restored from the original exercise copy when it has them, and any model-created ones are removed when it does not (conftest.py, pytest.ini, pyproject.toml, setup.cfg, tox.ini for Python, plus the equivalents for JS, C++, Java, Rust, Go).
• As a safety net, a 0 exit code is distrusted when the command's own summary reports failed tests (pytest's "N failed", mocha's "N failing", cargo's "test result: FAILED", go's "FAIL", gradle's "> Task FAILED" / "BUILD FAILED").

Testing

New tests in tests/test_benchmark_run_unit_tests.py:

• a model-created conftest.py can no longer force a pass: the exercise scores failed, and the conftest is removed before grading
• a pristine conftest.py is restored over a model-modified one
• an honest pass still passes
• the summary-pattern helper matches each runner's failure line and does not match zero-count summaries

Verified behaviorally in both directions with the exact scenario above: on the pre-fix code the exercise with a failing suite is recorded as solved via the conftest; with the change the conftest is removed and the exercise scores failed.

[CLAassistant]

All committers have signed the CLA.