fix(ai-chat): stop 100% CPU freeze during streaming (#1205)

[datlechin]

Summary

Fixes #1205. The AI chat panel pegged the main thread at 100% CPU and froze the app whenever a response streamed in. The cascade per 150 ms flush: existing + text rebuilt the assistant turn's text block (O(n²) over the response), plainText was recomputed from blocks on every observer tick, and the result fed into Markdown(text) keyed by \.offset, which made MarkdownUI re-parse the full markdown tree every flush. Schema prefetch and per-turn attachment resolution also ran on the main actor.

• Token append moved off the message blocks into an @ObservationIgnored buffer on the view model, published through a streamingTick counter so only the streaming bubble re-renders per token
• plainText is now a cached stored property maintained by appendText, not recomputed from blocks
• ChatContentBlock is a struct with a stable UUID id; the message-view ForEach keys on it, so finalized blocks are not re-created when a sibling changes
• The streaming bubble is its own subview (AIStreamingBubbleView) and renders plain Text; Markdown only runs once a block is committed at end-of-stream or tool boundary
• The redundant onChange(of: messages.last?.plainText) scroll trigger is gone; .defaultScrollAnchor(.bottom) + the messages.count trigger handle auto-scroll
• prepTask is now Task.detached(priority: .userInitiated); only hops to MainActor for capturePromptContext and messages.dropLast() snapshots

CHANGELOG entry added under Fixed.

Test plan

• Build green (xcodebuild, Debug)
• SwiftLint --strict clean on touched files
• AIChatViewModelMentionsTests passes (9/9)
• AnthropicProviderParserTests, GeminiProviderParserTests, OpenAICompatibleProviderParserTests pass
• Manual: send a short prompt with Include schema on, confirm response streams without UI hang and final text renders as Markdown
• Manual: send a tool-using prompt in Agent mode, confirm the assistant turn shows tool blocks + final text after each roundtrip
• Manual: cancel mid-stream, confirm partial text is preserved on the assistant turn and Stop ends streaming cleanly
• Manual: trigger a streaming error (revoke API key mid-stream), confirm error banner shows and empty turn is removed

Known follow-up

There is a separate UX issue where, after an Agent-mode tool roundtrip, a fresh assistant turn can be left with empty blocks if the model produces no text after the tool call — the message then sticks on the typing indicator. That is pre-existing (same path existed before this refactor) and is out of scope for this PR; it deserves its own fix to either prune the empty turn or render an explicit "completed" state.