Integrate LitePT and HEPT

[jpata]

This PR introduces major architectural updates, new model backends, and streamlined evaluation workflows.

• New Architectures: Integrated HEPT and LitePT models; refactored MLPF to support modular attention backends (FlashAttention, SDPA).
• Inference & Deployment: Improved ONNX export and expanded benchmarking scripts for CPU/GPU performance and memory profiling.
• Evaluation & Validation: Added a standalone Key4HEP evaluator for ROOT files (EDM4hep) and updated the ONNX validation suite for systematic model comparisons.
• Infrastructure & Tooling: added Hugging Face Hub integration scripts and a new detector hit visualization tool.
• Documentation: Updated README with streamlined guides for installation, dataset handling, and model evaluation.

LitePT is a lightweight, high-performance 3D point cloud architecture. LitePT-S has 3.6 × fewer parameters, 2 × faster runtime and 2 × lower memory footprint than PTv3. Moreover, it remains faster and more memory-efficient even when scaled up to LitePT-L with a parameter count twice that of PTv3. Already the smallest variant, LitePT-S, matches or outperforms state-of-the-art point cloud backbones across a range of benchmarks.

Valid loss: HEPT noticeably worse.

Jet match fraction: attention outperforms all.

Jet IQR: about the same for all models.

ONNX export check on GPU, attention has the best loss and GPU performance:

ONNX on CPU:

NB: A major caveat of this PR right now is that LitePT only supports Nvidia out of the box due to complex dependencies such as spconv-cu124. It also requires pinning torch to torch==2.9.0.