Allow users to write their own system prompt

[kisamoto]

System prompts change how the models respond and react. We currently support changing this via Agents but users should be able to add their own custom system prompts/agents to a library of them.

These are private and stored encrypted in the backend.

This is an extension of #65 and a stepping stone to allowing users to publish agents in the agent marketplace.