Add a REPL to:

• manage the main API server and lets the user discover and handle worker nodes
• manage local model storage and download of restricted models from HF
• trace callstack and get performance information at different levels (including interpreter-level with sys.setprofile)
• trace different subsystems (prefill, compute, prefetch) and specific components (tokenizer, attention, etc.)
• isolate and benchmark different subsystems and components with dummy data