feat(runtime): add Thinking pet state with 30s idleAfter for LLM processing gap

[JJ-zhang-92]

Problem

AI agent tools complete in ~1s but the LLM takes 5-20s to process results and generate text. During this gap, CoPet maps the "thinking" event to none (dropped) �� the pet transitions to Idle, leaving users to wonder if the agent crashed.

Changes (6 files, 20 lines)

Rust runtime

• Add PetStateId::Thinking variant to the enum
• Map "thinking" event kind to Thinking state
• Protect Thinking from being immediately overridden by Idle requests
• Give Thinking a 30s idleAfter (vs the default 1.5s for other temp states) �� LLM processing takes seconds, not milliseconds

Frontend

• Add "thinking" to PetStateId type union
• Map thinking AgentState in useAgentState.ts
• Route thinking to the review sprite row (Row 8) in petAnimation.ts
• Add thinking case to usePetSounds.ts

Why 30s instead of 1.5s

Upstream uses TEMP_STATE_IDLE_AFTER_MS = 1_500 for all temporary pet states. That works for Running/Review/Jumping �� these are transient tool visualizations. But Thinking represents an LLM processing gap that lasts 5-20s. A 1.5s timeout would cause the pet to snap into Idle while the agent is still working, defeating the purpose of the state.

Backward compatibility

• New enum variant only (no existing variants changed)
• Non-Thinking states still use 1.5s idleAfter
• "thinking" events from existing integrations are a no-op (they map to None today)

Tested

• Sending {"kind":"thinking"} via local runtime endpoint �� CoPet returns state: thinking with idleAfterMs ~30s
• tool.before events still get 1.5s idleAfter (unchanged)
• Thinking to idle transition after 30s confirmed