Cache a benchmark reference by checksum + hardware, measure on miss

When an autograder scores students as ratios against a reference implementation that it runs live on every submission, cache the reference's measured numbers instead. Key the cache by a checksum of everything that determines them (reference sources, benchmark/driver sources, config) plus the node's hardware identity (CPU model, core count, kernel). On a miss, run the reference live, score against it and save it, so the first submission after a staff change re-measures automatically, on the real grading hardware. Store the cache in a node-level dir mounted into the container, because per-job workspaces are wiped. Only trusted grader code may write it, behind a lock file and an atomic rename, and a corrupt entry is treated as a miss. Also recheck any worst-case runtime estimate against the per-run timeout formula: here 'every case hits the 120 s cap' gave 75 min, while the real min(120, 10x ref) worst case was about 15.