Add analysis doc for HurumNet vs notebook output discrepancies

[mtauraso]

Motivation

• Document likely causes for differing outputs between the self-contained notebook model and the Hyrax-converted model to aid debugging and reproduction efforts.
• Summarize architectural, preprocessing, and training/interpretation mismatches that can change logits or probabilities and prioritize fixes.
• Provide a clear, minimal reproduction procedure and actionable recommendations for aligning Hyrax behavior with the notebook reference.

Description

• Add a new analysis document at docs/hurum_net_discrepancy_analysis.md that enumerates observed mismatches and their consequences.
• Identify key discrepancies including an input-channel mismatch (KBModNet instantiated with in_channels=3 in the notebook vs hard-coded in_channels=1 in hurum_net.py), differing preprocessing pipelines (normalize_stamps() vs Hyrax percentile clipping/min-subtraction/sum normalization), and output-interpretation differences (notebook uses F.softmax(... )[:, 1] while Hyrax returns raw logits).
• Cover additional differences in training behavior such as target encoding, lack of gradient clipping, optimizer/scheduler setup, and class-balanced sampling, and recommend concrete checks and fixes.
• Provide a suggested minimal reproduction procedure and a prioritized fix order to resolve the most likely root causes first.

Testing

• No automated tests were added or run for this change because it is documentation-only.
• This change does not modify runtime code and therefore requires no code-based test updates to validate the analysis document.

---

Codex Task