load the system prompt on-the-fly instead of once at startup

this lets us modify it [for new conversations] on disk while the llm server
is running

server/inference.py

@@ -4,7 +4,6 @@ from .tid import tid_now
 from json import dumps as json
 
 db = sqlite_utils.Database(llm.cli.logs_db_path())
-girlypop_prompt = llm.cli.load_template("girlypop").system
 
 
 async def list_conversations(request: Request):
@@ -38,7 +37,7 @@ async def connect_to_conversation(ws: WebSocket):
   await ws.accept()
 
   # only send the system prompt at the start of a conversation
-  system_prompt = girlypop_prompt
+  system_prompt = llm.cli.load_template("girlypop").system
 
   if not continuing:
     for response in conversation.responses: