Add Interpretable Transformer Hawkes Process (iTHP)

[andrewwarrington]

Summary

• Add iTHP as an EasyTPP model, with package registration and a Taobao example configuration.
• Implement the active architecture of the authors' public code, adapted to EasyTPP event indexing, padding, likelihood evaluation, and sampled-time intensities. The model docstring cross-references the KDD 2024 paper and pinned public source files, recording where they disagree and why particular EasyTPP choices were made.
• The authors' active attention accepts but does not use n_head; this model likewise has one attention map. Monte Carlo integration draws random times within each observed interval by default. The authors' active script instead uses a batch-wide 0.1 grid. Trapezoid and per-interval fixed-grid modes are available explicitly. The auxiliary mark loss applies only during training.

Verification

• python -m pytest tests/ -q: 28 passed.
• Standard EasyTPP runner completed a one-epoch CPU smoke run on a small local fixture.

Scope and training settings

No shared EasyTPP runner, config, or evaluation code changes are included. The example preserves the selected model width, learning rate, and epoch budget. Our reported training runs used Adam epsilon 1e-5 and gradient-norm clipping at 1; stock EasyTPP does not apply those settings.