perf: concurrent schema fetch, memoize highlighting, cap tables

[datlechin]

Summary

3 performance fixes for the AI system — the final phase of the AI audit.

P2: Concurrent schema fetch with TaskGroup

• Replaced sequential for table in tablesToFetch loop with withTaskGroup
• Each table's column fetch (cache check + DB fallback) now runs concurrently
• 20 tables that previously took 1-4 seconds (sequential DB calls) now complete in ~200ms
• schemaProvider (actor) and driver are both safe for concurrent access

P4: Memoize code block syntax highlighting

• Added @State cache for highlighted AttributedString in AIChatCodeBlockView
• During streaming, re-highlights only when code grows by 50+ characters (vs every token)
• Eliminates O(n²) total work from per-token full regex passes
• currentHighlight() helper returns cached result when delta is small

P9b: Cap tables before fetch in SQLSchemaProvider

• SQLSchemaProvider.buildSchemaContextForAI() now caps to maxSchemaTables BEFORE the column fetch loop
• Previously iterated ALL tables (potentially 5000+), then buildSchemaSection() capped to 20 — wasting thousands of DB calls
• One-line fix: Array(tables.prefix(settings.maxSchemaTables))

Test plan

• Build succeeds
• Connect to a database with 20+ tables, open AI chat — schema loads quickly
• Send a message that produces a long SQL code block — no UI stutter during streaming
• Code blocks are properly syntax-highlighted after streaming completes
• Connect to a database with 100+ tables — AI chat opens without delay