StreamingVLA: Streaming Vision-Language-Action Model with Action Flow Matching and Adaptive Early Observation