On the Sample Complexity and Metastability of Heavy-tailed Policy Search in Continuous Control