Dump-overhead benchmarks: a dumper can cost more on the sim thread yet less wall and total CPU

When comparing a custom waveform capture against a built-in dumper, report three overheads per cycle against a no-dump run of the same executable: sim-thread CPU, all-thread CPU, and wall. They can disagree. In one small design the custom capture added more sim-thread CPU than the built-in dumper (1.62 vs 1.00 us/cycle), because of per-change public-interface calls, but less total CPU (7.0 vs 11.5) and less wall (2.6 vs 4.7), because its off-thread compression was cheaper and the built-in dumper's work held up the simulation. Use slopes from runs at several lengths: the built-in dumper had a large fixed startup cost that made the custom capture look better on the sim thread at 1M cycles than its real slope showed. State which of the three metrics a claim like 'faster than X' refers to.