Profiling capture callbacks: add one cost per arm to see which step dominates

To find where per-change capture overhead goes in a simulator callback path, build one executable with selectable arms that each add exactly one step: off (nothing registered), empty callback, copy the value, query sim time, and copy plus time. Then divide each arm's added CPU over off by the number of notifications. In one such breakdown the empty dispatch cost ~14 ns per change and copying the value added only ~5 ns. Querying the current simulation time on every change added ~25 ns, more than the dispatch itself, and copy+time was no higher than time alone. Many changes share a timestep, so fetch the time once per timestep (or on a time-advance callback) and reuse it, rather than per change. Report ns per notification and overhead against the no-capture run, and state the run count, because adjacent arms can differ by less than the noise.