Leaderboard grading: one grader test per benchmark case, metrics as result extras

Pattern for wiring a performance contest into a per-test autograder whose results feed a leaderboard uploader: generate one Tester class per contest case at registration time by parsing the scorer's own case list (so the two can't drift), and in each test run the scorer restricted to that case with a machine-readable output line (e.g. result NAME THREADS ok T_S P_NS B or failed REASON) rather than scraping human output, whose rounding loses precision. Return the raw metrics as result extras with exactly the keys the uploader maps, and on any failure still return the same keys zeroed so every submission has every case in the same order. Size each test's timeout from the scorer's per-run limit times runs plus build slack. Leave scoring to the leaderboard if it already scores against a benchmark submission.