Increase throughput for HLT2

The HLT2 throughput in 2025 is lower than desired. Listed are a few places where a speedup should be investigated. The most representative numbers / flame graphs are for the test MooreOnline_hlt2_pp_2025_data.

Some ideas:

• The selection line Hlt2QEE__JetBuilderGetSelJetsSimpleJets using the FastJetBuilder takes a considerable amount of time.
  • Mitigated in Rec!4408
• The evaluate function of SigmaNet takes more than 2% of the HLT2 sequence.
  • Mitigated in Rec!4395 (merged)
• Almost 3% of the full HLT2 sequence is used for Converters
• Investigate use of sorting and / or SIMD in PrLongLivedTracking
  • Mitigated in Rec!4427 (merged)
• Investigate timing of ParticleRangeFilter
• Neutral ProtoParticle making can be a lot faster
• Tighten ghost prob in accordance with better alignment (especially compared to running same tests (with 2025 software) on 2024 data)
• Retrain ghost prob with sim10f (when it arrives)
• Use LTO for compile flags, see e.g. here
• Use Ancestors to kill obvious clones (e.g. T track <-> Downstream, Velo <-> Upstream) in PrCloneKiller
  • Mitigated in: Rec!4464 (merged)
• ...