Compute sequence perplexity from token IDs and precomputed next-token logits using stable log probabilities.
