fix: resolve 4 critical AI bugs

[datlechin]

Summary

Fixes 4 ship-blocking bugs found during AI system audit:

• B1 — Ghost text at wrong offset: Inline suggestion ghost text rendered at position 0 (or previous suggestion's offset) during progressive streaming. suggestionOffset was set after fetchSuggestion() returned, but showGhostText was called during streaming using the stale value. Fix: set offset before the await.
• B2 — Tab key race condition: Pressing Tab to accept a suggestion did nothing because acceptSuggestion() was deferred via Task { @MainActor } to the next runloop tick, but handleTextChange() fired synchronously first and cleared currentSuggestion. Fix: call acceptSuggestion() synchronously (already inside MainActor.assumeIsolated).
• B3 — Trim breaks Anthropic API: trimMessagesIfNeeded() could leave the conversation starting with an assistant message after removing oldest messages, causing Anthropic API 400 errors ("messages must start with user role"). Fix: drop leading assistant messages after trim.
• B4 — Storage thread safety: AIChatStorage was a plain final class with no isolation. loadAll() ran on a detached thread while save()/delete() ran on MainActor — concurrent file I/O on the same directory. Fix: convert to actor, update all callers to await, rewrite deleteAll() to enumerate files instead of removing the directory.

Test plan

• Build succeeds with no new warnings
• Enable inline suggestions, type SQL — ghost text appears at cursor position (not at line start)
• Wait for ghost text, press Tab — suggestion inserts correctly
• Send 200+ messages in AI chat — no API errors from role ordering
• Open AI chat and send messages rapidly during app startup — no crashes or data loss