Use per-particle pt weight in loss

[jpata]

• enable sqrt(pt) weight term in loss that improves jet metrics
• switch CMS evals to v2.2.0/pyg-cms_20241212_101648_120237/mlpf_5M_attn2x3x245_bm5_relu_checkpoint5_1xa100_fp32_fused, trained with pt weight on 1xA100 for 5 epochs
  • CMSSW evals switched to NANOAOD (JME+BTV)
  • use JME NANO in CMSSW eval plots (notebooks/cms/cmssw-validation.ipynb)
  • add additional CMSSW validation datasets: QCD_noPU, ttbar_noPU, SingleEle, SinglePi, SingleGamma
  • update ONNX timing script
• also retrained CLIC model with the updated loss: https://huggingface.co/jpata/particleflow/blob/main/clic/clusters/v2.2.0/pyg-clic_20250106_193536_269746/README.md

Validation results:

• CMS, previous
  

• CMS, current
  

• CLIC, previous
  

• CLIC, current