FLUX.2-klein: CUDA OOM when running in for loop with different image

### Describe the bug

When using Flux2KleinPipeline for inference, running a single inference works fine and uses ~14GB VRAM. However, when running the exact same code inside a for loop, it immediately causes CUDA Out of Memory error on the first iteration, trying to allocate additional memory.
The OOM occurs specifically in apply_rotary_emb during the transformer's attention computation:
```
File "/venv/FluxApp/lib/python3.11/site-packages/diffusers/models/embeddings.py", line 1232, in apply_rotary_emb
out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 936.00 MiB. GPU 0 has a total capacity of 15.48 GiB of which 227.25 MiB is free. Including non-PyTorch memory, this process has 15.25 GiB memory in use. Of the allocated memory 14.78 GiB is allocated by PyTorch, and 148.94 MiB is reserved by PyTorch but unallocated.
```

### Reproduction

### Working code (single inference, ~14GB VRAM):

```python
ckpt_path = (
"https://huggingface.co/unsloth/FLUX.2-klein-4B-GGUF/blob/main/flux-2-klein-4b-Q6_K.gguf"
)

transformer = Flux2Transformer2DModel.from_single_file(
ckpt_path,
quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
torch_dtype=torch.bfloat16,
config="black-forest-labs/FLUX.2-klein-4B",
subfolder="transformer"
)

pipeline = Flux2KleinPipeline.from_pretrained(
"black-forest-labs/FLUX.2-klein-4B",
transformer=transformer,
torch_dtype=dtype
).to(device)

image = pipeline(
prompt="anime style",
image=image,
height=image.height,
width=image.width,
guidance_scale=1.0,
num_inference_steps=4,
).images[0]
```
### Failing code (for loop, OOM on first iteration):
```python
ckpt_path = (
"https://huggingface.co/unsloth/FLUX.2-klein-4B-GGUF/blob/main/flux-2-klein-4b-Q6_K.gguf"
)

transformer = Flux2Transformer2DModel.from_single_file(
ckpt_path,
quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
torch_dtype=torch.bfloat16,
config="black-forest-labs/FLUX.2-klein-4B",
subfolder="transformer"
)

pipeline = Flux2KleinPipeline.from_pretrained(
"black-forest-labs/FLUX.2-klein-4B",
transformer=transformer,
torch_dtype=dtype
).to(device)
for image_file in image_files:
base_image_obj = Image.open(image_file).convert("RGB")
image = pipeline(
prompt="anime style",
image=base_image_obj,
height=base_image_obj.height,
width=base_image_obj.width,
guidance_scale=1.0,
num_inference_steps=4,
).images[0]
```

### Logs

```shell
image = pipeline(
^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/diffusers/pipelines/flux2/pipeline_flux2_klein.py", line 843, in __call__
noise_pred = self.transformer(
^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/diffusers/models/transformers/transformer_flux2.py", line 874, in forward
encoder_hidden_states, hidden_states = block(
^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/diffusers/models/transformers/transformer_flux2.py", line 521, in forward
attention_outputs = self.attn(
^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/diffusers/models/transformers/transformer_flux2.py", line 256, in forward
return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, image_rotary_emb, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/diffusers/models/transformers/transformer_flux2.py", line 155, in __call__
query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/venv/FluxApp/lib/python3.11/site-packages/diffusers/models/embeddings.py", line 1232, in apply_rotary_emb
out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
~~~~~~~~~~^~~~~
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 936.00 MiB. GPU 0 has a total capacity of 15.48 GiB of which 227.25 MiB is free. Including non-PyTorch memory, this process has 15.25 GiB memory in use. Of the allocated memory 14.78 GiB is allocated by PyTorch, and 148.94 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
```

### System Info

- **OS**: Ubuntu 24.04
- **Python**: 3.11
- **PyTorch**: 2.9 (please specify your version)
- **diffusers**: latest (main branch)
- **GPU**: 16GB VRAM (RTX 5080)
- **CUDA**: 12.8

### Who can help?

@DN6

Reproduce the single-inference and loop cases using Flux2KleinPipeline, then trace memory use through diffusers/pipelines/flux2/pipeline_flux2_klein.py and diffusers/models/transformers/transformer_flux2.py. Inspect the apply_rotary_emb call in diffusers/models/embeddings.py and compare allocations between iterations. Done means identifying the loop-specific memory behavior and confirming that repeated inference no longer causes the reported CUDA OOM.

Scritto dal modello di indicizzazione a partire dal testo della issue.