Investigating Safety Vulnerabilities of Large Audio-Language Models Under Speaker Emotional Variations