fix: prevent --cpu flag from allocating GPU memory

Two root causes fixed:

1. soft_empty_cache() and synchronize() in model_management.py lacked a
   cpu_state == CPUState.CPU guard. They fell through to torch.cuda calls
   that initialize a CUDA context (150-500MB VRAM) even in CPU-only mode.

2. comfy_kitchen is imported unconditionally at startup via quant_ops.py.
   The import chain triggers torch.cuda.is_available() -> cuInit, which
   initializes the CUDA driver. Now gated behind args.cpu check.

Also adds missing QuantizedLayout and register_layout_op fallback stubs
that were absent from the original ImportError handler.

Amp-Thread-ID: https://ampcode.com/threads/T-019cbd03-433e-7601-93ff-3887227496b4

comfy/quant_ops.py

@@ -1,7 +1,11 @@
 import torch
 import logging
+from comfy.cli_args import args
 
-try:
+if args.cpu:
+    _CK_AVAILABLE = False
+else:
+    try:
         import comfy_kitchen as ck
         from comfy_kitchen.tensor import (
             QuantizedTensor,
@@ -24,10 +28,11 @@ try:
         ck.registry.disable("triton")
         for k, v in ck.list_backends().items():
             logging.info(f"Found comfy_kitchen backend {k}: {v}")
-except ImportError as e:
+    except ImportError as e:
         logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.")
         _CK_AVAILABLE = False
 
+if not _CK_AVAILABLE:
     class QuantizedTensor:
         pass