Fix multiclass LinearSVC fallback paths (closes #256)

[oegedijk]

Problem

Issue #256 reports multiclass LinearSVC failing in dashboard/explainer flows because it typically exposes decision_function but not predict_proba. Existing code paths assumed predict_proba, which caused:

• hard failures in PDP/permutation paths (AttributeError)
• class-count mismatches in SHAP flows when patched probabilities were malformed

What Changed

• Added a shared probability fallback that:
  • uses predict_proba when valid
  • falls back to decision_function and converts scores to probabilities (sigmoid for binary, softmax for multiclass)
• Applied fallback consistently across:
  • ClassifierExplainer prediction and kernel-SHAP callback paths
  • PDP generation (get_pdp_df)
  • one-vs-all and safe scorer paths used by permutation importances
• Added validation for class-count/shape mismatches to prevent silent bad outputs.

Tests

Added regression tests for multiclass LinearSVC:

• kernel SHAP path
• PDP path
• permutation importances path

Added guard tests for multiclass logistic regression to ensure existing predict_proba behavior remains unchanged.

Release Notes

Updated RELEASE_NOTES.md under 0.5.8 with issue #256 fixes and coverage.

Validation Run

• uv run pytest -q tests/test_linear_model.py
• uv run ruff check explainerdashboard/explainer_methods.py explainerdashboard/explainers.py tests/test_linear_model.py

Closes #256