agent.execute should accept file attachments along with the instruction

[nileshtrivedi]

Most LLMs now accept multi-modal input (images, PDF, audio). Agent.execute should accept the same and pass over to the underlying LLM. This can help with cases where critical context is non-textual.