feat: NORMAL_VRAM + AIMDO + copy=False patch + kernel caching

Major unified memory optimization changes:

1. model_management.py: HIGH_VRAM → NORMAL_VRAM
   - GB10 unified memory: offloading to CPU doesn't save physical RAM
     (same pool), but NORMAL_VRAM allows per-layer partial loading when
     memory is tight instead of all-or-nothing OOM
   - text_encoder_offload_device() and vae_offload_device() now return
     CPU (allows ComfyUI to offload unused models)
   - intermediate_device() still returns GPU (VAE outputs must stay in
     CUDA allocator for honest memory tracking)
   - User can force HIGH_VRAM with --highvram if models fit

2. utils.py: copy=True → copy=False for tensor.to(device)
   - On GB10 unified memory, copy=True creates a full duplicate in both
     CPU and CUDA allocators simultaneously (ComfyUI issue #10896)
   - copy=False makes .to(device) a zero-copy device label change since
     both allocators draw from the same physical LPDDR5X
   - Halves model loading memory usage when --disable-mmap is set

3. Removed --disable-dynamic-vram from ComfyUI flags
   - Was preventing AIMDO (comfy_aimdo) from initializing
   - AIMDO now activates: VBAR-based page-level VRAM management at 32MB
     granularity instead of blunt .to(cpu) copies
   - Falls back to NORMAL_VRAM per-layer loading if AIMDO has issues

4. Added CUDA_CACHE_MAXSIZE=4294967296 (4GB kernel cache)
   - PTX→SASS kernel caching for sm_121 (GB10 Blackwell)
   - 3x speedup on subsequent runs reported by DGX Spark community

5. System: vm.swappiness reduced from 60 to 1
   - Swap thrashing on unified memory causes silent system freezes
   - Near-zero swappiness ensures clean OOM kills instead

patches/model_management.py

@@ -507,13 +507,22 @@ def _is_unified_memory():
 UNIFIED_MEMORY = _is_unified_memory()
 
 if UNIFIED_MEMORY:
-    # On unified memory, offloading to CPU is pointless (same physical chips)
-    # HIGH_VRAM keeps everything on GPU and skips the offload/onload cycle
+    # On unified memory, NORMAL_VRAM allows ComfyUI to offload unused model
+    # layers to CPU when memory is tight. Since CPU and GPU share the same
+    # physical RAM on GB10, offloaded layers stay in the same physical pool
+    # but through a different allocator. Per-layer partial loading (LowVramPatch)
+    # means only individual layers are copied on-demand, not whole models,
+    # keeping peak memory manageable.
+    # HIGH_VRAM is available via --highvram if everything fits in VRAM.
     if not (args.highvram or args.gpu_only):
         logging.info("[Sparky] Grace-Blackwell unified memory detected — "
-                      "setting HIGH_VRAM mode (no CPU offloading)")
-    vram_state = VRAMState.HIGH_VRAM
-    logging.info(f"[Sparky] Set vram state to: {vram_state.name} (unified memory override)")
+                      "keeping NORMAL_VRAM mode (allows layer offloading)")
+    else:
+        logging.info("[Sparky] Grace-Blackwell unified memory detected — "
+                      "HIGH_VRAM requested via --highvram")
+    # Don't override vram_state — let ComfyUI's default NORMAL_VRAM handle
+    # offloading. User can force HIGH_VRAM with --highvram if models fit.
+    logging.info(f"[Sparky] Set vram state to: {vram_state.name} (unified memory)")
 else:
     logging.info(f"Set vram state to: {vram_state.name}")
 
@@ -1054,7 +1063,7 @@ def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.flo
     return torch.float32
 
 def text_encoder_offload_device():
-    if args.gpu_only or UNIFIED_MEMORY:
+    if args.gpu_only:
         return get_torch_device()
     else:
         return torch.device("cpu")
@@ -1123,7 +1132,7 @@ def vae_device():
     return get_torch_device()
 
 def vae_offload_device():
-    if args.gpu_only or UNIFIED_MEMORY:
+    if args.gpu_only:
         return get_torch_device()
     else:
         return torch.device("cpu")