VPI change capture: notify-then-read-back loses to copying the callback's own value

When measuring ways to capture RTL signal changes in a commercial simulator (one existing simv, about 9M changes at 25K cycles, n=3), the cheapest route was to copy the value already delivered with the change event. Pair that with a ring buffer, a cached time value and adaptive block sizes. A dirty-notification route that then fetched each settled value with read-only VPI calls cost about 27% more CPU and 43% more wall time. Pure public VPI value-change callbacks were worse still. Notifications only cut delivered records by about 2%, because nearly every notified signal still had to be read. So don't assume 'notify, then read on demand' saves work for dense designs. It only helps when few notified signals are ever read. Caching simulation time instead of querying it per event also mattered.