feat(mecatui): surface live LLM stream progress

Problem

During a long LLM call, mecatui can appear stalled even while the server is receiving meaningful stream output. The LLM idle watchdog is reset by adapter-level chunks, but the UI should make that progress visible rather than leaving users with only an indeterminate wait.

Desired behaviour

While a turn is in progress, render observable progress for every meaningful provider chunk that reaches the client:

• streamed assistant text/tokens;
• reasoning/thinking updates; and
• tool-call progress, including the completed tool call when it becomes available.

The presentation should update incrementally without duplicating final transcript/tool-card content, preserve scrollback usability, and work for both embedded mecatui and mecatui connect sessions.

Acceptance criteria

1. A streamed text chunk produces visible incremental assistant progress before turn completion.
2. A reasoning/thinking chunk produces visible progress while reasoning is streaming.
3. Tool-call progress is visibly represented when the server emits it, without duplicating the final tool card.
4. Progress is cleared, finalized, or folded into the normal transcript at terminal turn completion; it must not leave stale in-progress UI behind.
5. Add focused UI/client tests covering text, reasoning, tool-call, and terminal cleanup behavior.

Context

The server-side idle watchdog resets when provider adapters yield port.Chunks, not on SSE keepalives. Text and reasoning deltas yield chunks; raw lifecycle events, keepalives, and partial tool-argument fragments may not. Showing the chunks that do arrive gives users evidence of active model progress during long calls.