Implement event batching

[ggalgoczi]

When handing photons (gensteps) from Geant4 to the GPU, we need a way to identify which G4 event each photon belongs to after GPU ray tracing completes.

Approach: Encode the event ID into the photon's hit time by adding eventID*M, where M is a large multiplier. After GPU ray tracing, the original event ID can be recovered from each hit via hitTime / M (integer division), and the actual hit time via hitTime % M.

Constraint: M must be larger than the maximum possible hit time (i.e., longer than the longest-lived particle's lifetime) to avoid collisions between the encoded event ID and the real time value.

[ggalgoczi]

This was done previously in #21 but gone now. Guess is it was overwritten with a sync with upstream? Anyway there is current max limit in the GPU code that kills photons but set to a very high number: 1.e27f ns.

https://github.com/BNLNPPS/eic-opticks/blob/main/sysrap/SEventConfig.cc#L54

[ggalgoczi]

Let's take into account the case of DUNE. From the literature Scintillation emission time constants: Triplet (slow component): ~1.3–1.6 µs. We have to take into account the nuclear decays though. 10 ms with TPB delayed emission. However if nuclear decays are modeled the photon time can be even longer. From what I see Ar-39 has half life of 269 years. @brettviren is there any longer lived relevant isotope? Anyway I guess would get killed in simulations because DUNE will not last hundreds of years : )

[ggalgoczi]

Investigated two approaches for mapping GPU hits back to G4 events when batching multiple events into a single GPU launch. The time-encoding approach (time_hack_event_batching) adds eventID * 1e6 ns to the genstep time in U4.cc:118 — this is simple (9 lines) but limited by float32 precision in sphoton.time (~7 significant digits), causing a few hits per event to be misassigned at boundaries, and breaks entirely for long-lived nuclear decays where parent times exceed the multiplier. The seed-buffer approach (event-batching) uses the existing GPU seed buffer (seed[photon_idx] → genstep_idx) combined with a CPU-side EventBatcher that records which gensteps belong to which G4 event — this is exact, has zero precision loss, and costs only ~0.16 ms to download the seed buffer for 1M photons. Also moved QEvt::gatherSeed() out of the #ifndef PRODUCTION guard so the seed buffer is accessible in production builds. Both branches include modified GPURaytrace.h that writes per-event hit files (opticks_hits_evt{N}.txt) with decoded times. The time-hack is a quick-and-dirty solution suitable for short-lived physics, while the seed-buffer approach is the correct long-term solution for arbitrary physics including long-lived isotopes.