Cancel in-flight LLM stream on thread switch or new chat

[theantichris]

Description

Running :n, loading a thread from :t, or pressing Escape in normal mode while the LLM is streaming can cause a race condition. The streaming goroutine continues sending LLMResponseMsg/LLMDoneMsg after the chat state has been reset, which can:

• Append stale response chunks to the new/loaded conversation's chatHistory
• Persist the assistant message under the wrong (new or empty) threadID
• Race on model.messages (goroutine writes on line 30 of chat_stream.go)

Proposed solution

Add a per-stream cancellable context and a streaming flag to ChatModel:

• streamCtx, streamCancel := context.WithCancel(model.ctx) created in startLLMStream()
• streaming bool set true in startLLMStream(), false in handleLLMDoneMsg/handleLLMErrorMsg/cancel handler
• A cancelStream() method that calls streamCancel(), clears currentResponse, and sets streaming = false

Cancel trigger points

• Escape in normal mode while streaming → cancel stream
• :n (newChat()) → cancel stream, then reset state
• loadThread() (selecting a thread from :t list) → cancel stream, then load thread

:t itself does not cancel — it just opens the list overlay. The stream can continue underneath until a thread is actually selected.

Post-cancel cleanup

• handleLLMResponseMsg should early-return if !model.streaming
• Distinguish intentional cancel from errors (e.g. LLMCancelledMsg vs LLMErrorMsg) so cancelled streams don't show error messages

Split from #247.