AI timeline groups tool calls for cleaner chat

The AI chat timeline could get noisy — multiple tool calls, intermediate thinking text, and subagent outputs all rendered inline made it hard to follow the conversation. Now response groups collapse these elements behind expandable summaries showing tool call counts and icons. Tool calls and subagent spawns are grouped together by their API response, with the actual answer text extracted and always visible outside the group.

The responseId tracking runs through the event pipeline. When the LLM stream emits a step-start chunk, the responseId is captured before any trace hooks run. Agent-spawned tool calls inherit the responseId from their parent's last entry since those events originate from tool code, not the stream executor. Text entries from the same response as an active group stay inside it as intermediate thinking text.

A memory leak in InstanceAiMarkdown was also fixed — click handlers on resource links were added on each render without cleanup, causing listeners to accumulate. A WeakMap now tracks handlers so they can be removed on unmount and update. The initial listener count dropped from roughly 7,900 to 4,900.

The collapsible animations use a new AnimatedCollapsibleContent component with slide-down and slide-up CSS animations, replacing the standard reka-ui CollapsibleContent across all collapsible sections. The PlanReviewPanel header became a collapsible trigger so the whole panel collapses after approval.

Visual refinements from the visual-tweaks branch standardized button sizes to medium, adjusted card backgrounds to use a consistent light-3 shade, improved dark mode contrast on answered questions, and refined typography in task details and JSON viewers.

This builds on #28155 and groups the instance AI timeline with response grouping and UI improvements.