InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents