Micro-benchmarks as regression tests for OCaml

Thumper measures OCaml micro-benchmarks and checks them against a committed baseline, so a dune rule fails the build when a case regresses beyond its budget. The baseline stores the evidence, sorted timing samples and exact allocation counts, in one section per machine. A timing comparison is a rank test, and a regression or improvement stands only when a confirmation pass in the same run reproduces it; an exact count compares as an integer. A comparison that cannot decide is inconclusive and fails nothing. A case with no baseline row yet is proposed and passes, and fails only under --strict; bless records the baseline. Thumper depends on the unix library alone. On Unix each case runs in a forked worker.