fix: retry NvAPI init on autoload to handle early boot race

init_nvapi() previously called sys.exit(1) directly, which raised
SystemExit (uncaught by except Exception). The autoload subprocess
would die instantly if the NVIDIA driver wasn't fully loaded at boot.

- Raise RuntimeError from init_nvapi() so callers can handle failures
- Add 16-attempt retry loop (2s interval, 32s total) in run_autoload()

nvcurve/hal/gpu.py

@@ -12,8 +12,7 @@ def init_nvapi() -> None:
     """Initialize NvAPI. Must be called before any GPU operations."""
     init_fn = query_interface(FUNC["Initialize"], nargs=0)
     if not init_fn or init_fn() != 0:
-        print("NvAPI_Initialize failed")
-        sys.exit(1)
+        raise RuntimeError("NvAPI_Initialize failed")
 
 
 def enumerate_gpus() -> tuple[ctypes.Array, int]: