MCKinematicsReader: keep the loaded tracks instead of copying them, free the baskets

loadTracksForSourceAndEvent deep-copied the vector which ROOT allocated for us (a pointer to nullptr was
passed, so it is ours) and deleted the original: +1x the event at peak (373 MB for a PbPb event with 6.5M
MCTracks). The vector is now stored directly and the branch address reset. In addition, the decompressed
baskets (~1x the event for the split MCTrack branch) are dropped once the event is the last entry of its
TTree cluster, i.e. when no later entry can reuse them, which keeps small events sharing baskets as fast
as before (QED, 10k events: 0.17 s either way, 0.70 s when dropping after every event). For the PbPb
event: peak memory in use 1628 -> 1232 MB, held after releaseTracksForSourceAndEvent 882 -> 443 MB.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: 2 people

Steer/src/MCKinematicsReader.cxx

@@ -92,9 +92,15 @@ void MCKinematicsReader::loadTracksForSourceAndEvent(int source, int event) cons
       std::vector<MCTrack>* loadtracks = nullptr;
       br->SetAddress(&loadtracks);
       br->GetEntry(event);
-      mTracks[source][event] = new std::vector<o2::MCTrack>;
-      *mTracks[source][event] = *loadtracks;
-      delete loadtracks;
+      // ROOT allocated the vector for us and we own it (we passed a pointer to nullptr): keep it instead of copying it
+      mTracks[source][event] = loadtracks;
+      br->ResetAddress(); // the branch must not refer to the stored vector (nor to the local pointer) any more
+      // free the decompressed baskets (~ the size of the event) if no later entry reads them, i.e. at the end of its cluster
+      auto clusterIt = br->GetTree()->GetClusterIterator(event);
+      clusterIt.Next();
+      if (event + 1 >= clusterIt.GetNextEntry()) {
+        br->DropBaskets("all");
+      }
     }
   }
 }