A Training Criterion with Token-Level Tolerance to Transcription Ambiguity for Automatic Speech Recognition — ThinkLLM