Abstraction for browser automation tasks

[rcholic]

Summary

Abstract WebBenchTask to a generic AutomationTask for the SDK's PlannerExecutorAgent to handle broad web automation tasks like "buy a laptop on xyz.com". The design removes hardcoded WebBench dependencies, enables dynamic heuristics composition by the planner, and implements rollback/recovery to last known good state.

---

Design Goals

1. Generic Task Model: Replace WebBench-specific WebBenchTask with AutomationTask that works for any browser automation
2. Dynamic Heuristics: Let planner compose element-selection heuristics on-the-fly via HeuristicHint
3. Deterministic Verification: Use existing PredicateSpec system (url_contains, exists, etc.) for post-step verification
4. Recovery/Rollback: Implement URL-based recovery when verification fails repeatedly
5. Backward Compatibility: WebBench can continue using WebBenchTask via factory method