HUME: Measuring the Human-Model Performance Gap in Text Embedding Tasks