feat: add three-layer agent loop protection

[Jonnytoshen]

Summary

Adds a three-layer agent loop protection system to prevent infinite loops, along with supporting infrastructure (mock model, utility types, and new dependency).

Changes

feat(types): SafeAny type

• Added SafeAny type alias to src/types/any.ts with a barrel export from src/types/index.ts.

feat: Mock model implementation

• src/mock-model.ts — a fully configurable in-process model stub that supports tool calling and scripted response sequences, enabling deterministic loop testing without live API calls.

feat: Three-layer agent loop protection

• Loop detection (src/loop-detection.ts): Tracks tool call history with a sliding window. Injects a warning message after 5 repeated calls and halts execution at 8 (critical threshold) or 10 (global circuit breaker).
• Step-level retry (src/retry.ts): Exponential backoff with up to MAX_RETRIES=3 for retryable API errors (rate limit, timeout, network failures).
• Token budget: Accumulates input + output tokens across steps and force-stops when the limit is exceeded (default 15,000 tokens).
• Updated src/agent-loop.ts and src/index.ts to integrate all three layers.

chore: Add @ai-sdk/provider dependency

• Required peer dependency for the provider abstraction used by the mock model.

Testing

• Lint: ✅ passed
• Format check: ✅ passed