Describir: Systematically Analysing Prompt Injection Vulnerabilities in Diverse LLM Architectures