This PR updates the nanobench with the latest release from upstream, v4.3.6. It fixes the missing performance counters. Due to discussions on #22999 I have done some work that should make the benchmark results more reliable. It introduces a new flag -min_time that allows to run a benchmark for much longer then the default. When results are unreliable, choosing a large timeframe here should usually get repeatable results even when frequency scaling cannot be disabled. The default is now 10ms. For this to work I have changed the AddrManGood and EvictionProtection benchmarks so they work with any number of iterations.
Backport of core#23025.
Depends on D12902.