Bugs: Fixing issue with emote encoding in streaming chat, Fixing issue with missing pad_token for pytorch tokenizers, allowing system message as latest message in chat

[AndiMajore]

Collection of three small fixes for issues:

• Streaming mode for chats broke emote encoding; fixed by trimming incomplete/non-decodable characters from chunk
• Running llama models would throw missing pad_token errors: fixed by setting to eos_token
• Ran into issue when initializing behavior by sending system message only: allowing system or user messages to be latest message

[aresnow1]

Could you provide a minimal reproduction for the first issue?

I've reproduced and tested on this branch, thanks for your contributions!