Fix GEMMA_2, LLAMA_3, and PHI_3 MessagesFormatter

Gemma needs another \n in the prompt template to avoid slow processing of follow up prompts, see #54  As can be seen in the attached screenshot the prompt processing is much faster. Before: The new user prompt plus the models previous answer are processed (42 tokens) After: Only the new user prompt is processed (18 tokens) The longer the models previous answer the more relevant... Probably there are similar issues in all other prompt templates.