Render scheduled wakeups as an inline tool line

[stippi]

What

schedule_wakeup gets its own unobtrusive inline rendering in the GPUI frontend, and the injected wakeup message no longer looks like a message from the user.

Tool line (alarm clock icon, Explore/Read style):

• before execution: "Schedule wakeup in 20 min"
• armed: "Wakeup at 14:31 · in 20 min" — live countdown, seconds in the last minute; "on Oct 13 at 09:00" for other days
• fired: "Woke up at 14:31" with the alarm-clock-check icon
• expanding it shows the prompt
• due but never fired (restart, cancel): neutral "Wakeup at 14:31"
• cancel_wakeup: "Cancel wakeup Add real UI based on Zed's gpui crate #3" with the alarm-clock-off icon

Wakeup message

• follows its schedule_wakeup block within the same turn (matched by prompt) → not shown; the block says it fired
• otherwise (the user spoke in between) → shown as a tool-like "Woke up" line with the prompt, no "You" card

How

• Core: ScheduleWakeupOutput carries fires_at; render_for_ui returns JSON (like spawn_agent), the LLM still sees the text. Old records without fires_at still parse. session::wakeup::wakeup_prompt recognizes injected messages.
• ACP: renders the JSON result as "Wakeup Add real UI based on Zed's gpui crate #3 armed, fires at …".
• GPUI: WakeupToolRenderer; ToolBlockRenderer gains icon() and refresh_after() (defaults keep existing behavior). Wakeup recognition runs on every path that creates user messages (live input, session load, append, branch edit).

Tests

• renderer descriptions and refresh timing
• main_screen::tests::wakeups: loaded session with a linked and an unlinked wakeup
• cargo test, clippy -D warnings, fmt --check green

Known gap: a wakeup that fires while the agent runs briefly shows as pending message with the [scheduled wakeup] prefix.