Variance-Reduced Conditional Gradient Methods under Markovian Sampling for Nonconvex Composite Optimization