docs: align data description and refresh example metrics

[AmirhosseinHonardoust]

Fixes inconsistencies between the docs and the actual workflow.

• Replace the inaccurate "highly separable" wording in train_model.py,
  dashboard_utils.py, and app.py. The generator deliberately produces
  overlapping classes with label noise, so the data is not highly
  separable. Wording now matches the README and generator.
• Refresh the README example-metrics table to current values
  reproducible with --seed 42 (ROC-AUC unchanged at 0.976; AP, Brier,
  threshold, precision, recall, and flagged rate updated). Add a note
  that exact values can vary across scikit-learn/NumPy versions.
• Add SciPy to the Tech Stack list (now an explicit dependency).

Verified locally: ruff/black/mypy clean, 34/34 tests pass (including
the data_note "synthetic" assertion). Metrics confirmed deterministic
across repeated runs.