[TEST PR] Script for CI cosine similarity comparisons

[leesharkey]

Description

Enhances the model comparison script with CI-based cosine similarity metrics for
more meaningful analysis of learned component alignment between SPD model runs.

Key Changes:

Metric Improvements:

• Replaced activation density with mean causal importance (CI) as the component
  filtering criterion
• Added CI cosine similarity metrics to measure component alignment between models
• Renamed density_threshold → mean_ci_threshold with proper validation (0.0-1.0
  range)

Code Quality:

• Refactored compute_activation_densities() → compute_ci_statistics() with better
  error handling
• Added comprehensive shape mismatch checking with detailed warnings
• Improved batch handling with StopIteration safeguards
• Enhanced logging with component-level statistics

Configuration Updates:

• Updated compare_models_config.yaml with new semantic parameter names
• Adjusted default threshold value for CI-based filtering
• Updated example model paths and batch size

Related Issue

N/A - Enhancement to post-hoc analysis tooling

Motivation and Context

The CI cosine similarity metrics provide better insight into how learned components
align between different model runs. Mean CI is a more meaningful measure of
component importance than activation density, as it directly quantifies each
component's causal contribution to model outputs.

This complements the existing geometric similarity metrics (which compare component
subspace geometry) with a functional similarity metric (which compares component
usage patterns on actual data).

How Has This Been Tested?

• ✅ All formatting checks pass (make check)
• ✅ All type checks pass (basedpyright, 0 errors)
• ✅ All unit tests pass (200 passed, 11 skipped)
• ✅ Code reviewed against CLAUDE_CHECKLIST.md standards
• ✅ Removed obvious comment per style guide

Does this PR introduce a breaking change?

Minor breaking change in compare_models.py:

• Config parameter renamed: density_threshold → mean_ci_threshold
• Users with existing compare_models_config.yaml files will need to update this
  parameter name
• Impact is minimal: The script is for post-hoc analysis only, not part of the core
  SPD training pipeline