Avoiding Inference Heuristics in Few-shot Prompt-based Finetuning