Improve betweenness centrality for large codebases

[lambda-alpha-labs]

Brandes' algorithm in src/analyze/questions.rs is capped at 5,000 nodes. For larger codebases, key bottlenecks in later modules may be missed.

Ideas

• Run centrality per Louvain community instead of globally
• Implement an approximate algorithm (k-path centrality or random walk) for graphs exceeding the cap
• Only recompute changed communities during incremental runs
• Show progress during centrality computation

Acceptance

• Centrality analysis scales beyond 5,000 nodes without silently truncating
• Minimal performance regression on small graphs