Python 3.12: Torch CUDA extensions fail to import unless torch DLL directory is registered

## Python 3.12: Torch CUDA extensions fail to import unless torch DLL directory is registered

### Description

When using **Python 3.12 + PyTorch 2.10 (CUDA 13)**, compiled CUDA extensions fail to load with:

```
ImportError: DLL load failed while importing : The specified module could not be found.
```

Example extension:

```
custom_rasterizer_kernel.cp312-win_amd64.pyd
```

The extension compiles successfully but fails to import.

---

### Environment

```
Windows 11
Python 3.12
Torch 2.10.0+cu130
CUDA 13
ComfyUI: 0.16.4
comfy-aimdo: 0.2.9
comfy-kitchen: 0.2.7
ComfyUI frontend: 1.41.11
```

Torch runtime libraries exist correctly in:

```
Python312/Lib/site-packages/torch/lib
```

Example files:

```
c10.dll
c10_cuda.dll
torch_cpu.dll
torch_cuda.dll
torch_python.dll
```

---

### Cause

Python ≥3.8 tightened DLL loading behavior and does **not search PATH automatically** for extension dependencies.

CUDA extensions that depend on Torch runtime libraries therefore fail to load unless the Torch DLL directory is explicitly registered.

---

### Workaround

Registering the Torch DLL directory before importing the extension fixes the issue:

```python
import os
os.add_dll_directory(r".../site-packages/torch/lib")

import torch
import custom_rasterizer_kernel
```

This allows the extension to load successfully.

---

### Suggested Fix

ComfyUI could register the Torch DLL directory before importing torch:

```python
import os, torch
os.add_dll_directory(os.path.join(os.path.dirname(torch.__file__), "lib"))
```

This would ensure that **all Torch CUDA extensions load correctly**, including those from custom nodes.

---

### Why this matters

New nodes using CUDA extensions (e.g. rasterizers, Triton kernels, 3D nodes) are increasingly common and may hit this issue with:

```
Python 3.12
Torch 2.10+
CUDA 13
```

Registering the Torch DLL directory once during startup would prevent these import errors.