Incremental Policy Iteration with Guaranteed Escape from Local Optima in POMDP Planning