Clone the registry GTouchable per step to stop cross-event state leak

[zhaozhiwen]

getGTouchable returned the persistent shared_ptr<GTouchable> stored in gTouchableMap, and ProcessHits/processTouchable mutate it each step (trackId, pid, stepTimeAtElectronicsIndex). Initialize() clears the per-event touchableVector but never resets those mutations, so at the start of each event the registry object still holds the previous event's time-cell index. processTouchableImpl branches on that index, so the first step of a new event can be misrouted (a spurious clone / wrong time cell), and stale trackId/pid carry over.

Return a per-step copy so the registry entry stays pristine. Each step then starts from the clean UNSET identity, and within-event grouping still works because the hit collection / touchableVector group by operator== over field values (freshly assigned per step), not by object identity. gsd.h documents instances as thread-local with a per-worker map, so the clone introduces no cross-thread sharing.

Validation: ran the cherenkov example (gPhotonDetector SD, 20 events, fixed -seed) before and after — the hit physics is byte-identical (only the wall-clock timestamp column differs), i.e. no regression, while the fix removes the cross-event index carryover that affects readout SDs (Geant4 11.4.1 dev container).

Note: this is the issue's primary (clone) fix. For readout-type SDs it changes behavior by design — it stops the cross-event bleed and the spurious double-cell contribution from processTouchableImpl's clone-on-change branch; each step now routes to exactly one time cell via operator==.

Fixes #113