Fix cluster card conversation count mismatch

Backend:
- Add total_unique_conversations calculation in _compute_cluster_scores()
- Include total_unique_conversations in flatten_cluster_scores() output
- Correctly count distinct conversations per cluster using conversation_id.nunique()

Frontend (submodule):
- Add clusterScores prop to ClustersTab
- Update enrichment logic to read from cluster_scores
- Change labels: 'conversations' -> 'unique conversations', 'responses' in tooltips
- Pass cluster_scores from App.tsx to ClustersTab

This fixes the mismatch where cluster card totals didn't match the sum of
per-model bar counts. The root cause was that unique conversations (distinct
prompts) differs from model responses (one conversation can have multiple
model responses in multi-model datasets).

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Author: lisadunlap