refactor: clean AI architecture

[datlechin]

Summary

Architecture cleanup for the AI system — 7 improvements across 9 files, net -44 lines.

A5: Deduplicate provider code

• Extracted identical collectErrorBody() (9 lines × 3 providers) into AIProvider protocol extension
• Extracted mapHTTPError() base implementation into AIProviderError static method
• Only GeminiProvider retains a thin override (adds 403 → auth failure mapping)
• Removed ~55 lines of duplicated code

A4: Cache provider instances

• AIProviderFactory now caches providers keyed by (config.id, apiKey)
• Eliminates URLSession recreation on every inline suggestion request (~2x/sec)
• Added invalidateCache() and invalidateCache(for:) methods
• Cache invalidated from AISettingsView on provider save/delete

A3: Unified provider resolution

• Added AIProviderFactory.ResolvedProvider struct and resolve(for:settings:) convenience
• Replaces 3-step resolveProvider → resolveModel → createProvider pattern
• Updated both callers: InlineSuggestionManager and AIChatViewModel

A8: Use AIPromptTemplates

• handleFixError, handleExplainSelection, handleOptimizeSelection now use AIPromptTemplates instead of building prompts inline
• Removed dead queryLanguage and queryTypeName computed properties

A2: Move schema fetch to ViewModel

• Moved fetchSchemaContext() from AIChatPanelView (View) to AIChatViewModel
• Business logic no longer lives in a SwiftUI view struct
• View calls await viewModel.fetchSchemaContext() in .task(id:)

UI1: Use shared EmptyStateView

• Replaced ContentUnavailableView with app's EmptyStateView component
• Consistent styling via ThemeEngine across all panels
• Removed SettingsLink hack with raw UserDefaults key

UI2: Fix scroll-to-bottom

• Removed invisible Color.clear anchor element and isNearBottom state tracking
• Scroll targets use actual message IDs directly
• Removed DispatchQueue.main.async workarounds
• Single scrollToBottom(proxy:) helper replaces 4 duplicate scroll blocks

Test plan

• Build succeeds
• AI chat: send messages, streaming works, scroll follows output
• Inline suggestions: appear correctly, Tab accepts
• Explain/Optimize/Fix Error from context menu work
• Settings: add/edit/delete provider, changes take effect immediately
• Empty states display correctly when no provider configured
• Switch conversations, scroll resets to bottom