Earley-Driven Dynamic Pruning for Efficient Structured Decoding