v0.33.1版本更新导致flux2无法正常使用gguf格式的clip模型

v0.33.1 的 `comfy/text_encoders/llama.py` 中，8月13日 "Cuda Graphs" 相关提交将 TransformerBlock 的 MLP 后残差相加从：

```python
x = residual + x
```

改为：

```python
x = torch.add(residual, x, out=output) # 配合函数开头 output = x，原地写回 block 输入张量
```

该写法对普通 PyTorch 模型数学等价，但**实际改变了 Qwen3-8B GGUF CLIP 的编码输出数值**，导致FLUX2 条件编码异常 → 采样偏离提示词、图片过曝泛白。

临时修复方案

修改 `ComfyUI\comfy\text_encoders\llama.py`，将两处（TransformerBlock 与 TransformerBlockGemma2）的 `torch.add(residual, x, out=output)` 恢复为 `x = residual + x`，并删除无用的 `output = x`：

```python
# Self Attention
residual = x
x = self.input_layernorm(x)
x, present_key_value = self.self_attn(...)
x = residual + x

# MLP
residual = x
x = self.post_attention_layernorm(x)
x = self.mlp(x)
x = residual + x # 恢复：原为 torch.add(residual, x, out=output)
```

Start in comfy/text_encoders/llama.py and inspect TransformerBlock and TransformerBlockGemma2, especially the residual additions changed by the Cuda Graphs commit. Reproduce FLUX2 conditioning with a Qwen3-8B GGUF CLIP model, then verify that restoring the reported residual behavior prevents prompt deviation and overexposed output.

Written by the indexing model from the issue text.