[tamnd]

Plan 12.2 says a benchmark is within 5% when the upper end of the 95% interval of its rucc/gcc median ratio is at most 1.05, and K8's exit criterion says the same. rk bench only looked at the ratio of the two medians, so one lucky run could pass a benchmark that is really slower.

Bench::interval resamples the runs of each kernel with replacement 2000 times, takes the ratio of the medians each time and keeps the 2.5th and 97.5th percentiles. The seed is fixed, so the same bench.json always gives the same interval. The gate now checks the upper end of each interval instead of the ratio. The geometric mean check stays on the medians, as the exit criterion has it. The table gets a "95% interval" column.

New test: equal runs give an interval of just the ratio, a spread of a few percent widens it past 1.05, and an empty bench has none. cargo test -p rk bench and clippy pass on server3.

Part of #9.