dynamic_vram: Fix windows Aimdo crash + Fix LLM performance (#12408)

* model_management: lazy-cache aimdo_tensor

These tensors cosntructed from aimdo-allocations are CPU expensive to
make on the pytorch side. Add a cache version that will be valid with
signature match to fast path past whatever torch is doing.

* dynamic_vram: Minimize fast path CPU work

Move as much as possible inside the not resident if block and cache
the formed weight and bias rather than the flat intermediates. In
extreme layer weight rates this adds up.

comfy/model_patcher.py

@@ -1542,7 +1542,6 @@ class ModelPatcherDynamic(ModelPatcher):
 
                     if vbar is not None and not hasattr(m, "_v"):
                         m._v = vbar.alloc(v_weight_size)
-                        m._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(m._v, device_to)
                     allocated_size += v_weight_size
 
                 else:
@@ -1557,7 +1556,6 @@ class ModelPatcherDynamic(ModelPatcher):
                         weight_size = geometry.numel() * geometry.element_size()
                         if vbar is not None and not hasattr(weight, "_v"):
                             weight._v = vbar.alloc(weight_size)
-                            weight._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device_to)
                             weight._model_dtype = model_dtype
                         allocated_size += weight_size
                     vbar.set_watermark_limit(allocated_size)