FELLE: Autoregressive Speech Synthesis with Token-Wise Coarse-to-Fine Flow Matching