Dynamic-Width Speculative Beam Decoding for Efficient LLM Inference