Fix image_grid_thw to be in CPU

[cyyever]

What does this PR do?

In vLLm inference and SFT training, there are lots of blocking operations of image_grid_thw such as torch.prod and torch.tolist, so let's always fix image_grid_thw to CPU to avoid them.
A simple grep gives the following examples:

src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py:        split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
src/transformers/models/qwen2_vl/modeling_qwen2_vl.py:        split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py:                    samples = torch.split(image_grid_thw, list(image_nums))
src/transformers/models/qwen2_vl/modeling_qwen2_vl.py:                    samples = torch.split(image_grid_thw, list(image_nums))
src/transformers/models/glm4v/modeling_glm4v.py:                    samples = torch.split(image_grid_thw, list(image_nums))

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: glm4v, qwen2_vl

[Rocketknight1]

cc @qubvel @zucchini-nlp

[zucchini-nlp]

I don't think we need this. The model needs all inputs to be in the same device and users who pass in device='cuda' will expect image grid tensors to be moved to CUDA when processing. This leads to inconsistencies in API

I'd better have vLLM move the tensor back to CPU if needed or call image processor with device="cpu"

[qubvel]

We might need to provide pixel_values.device to avoid creating a tensor on CPU by default, wdyt?

[cyyever]

@zucchini-nlp @qubvel This tensor is special because its main purpose is index manipulation. For such tensors for indexing, moving them to GPU doesn't provide acceleration.

The dilemma is that we want almost all of tensors to be in CUDA except some special ones, so using device for processor is not a good solution.
Admittedly this is a semantics-breaking change..

[zucchini-nlp]

This tensor is special because its main purpose is index manipulation. For such tensors for indexing, moving them to GPU doesn't provide acceleration.

I see, though it will be quite breaking and as noted above, will not be consistent with what image processing API does. IMO all inputs returned will need to be in the same device and type, as requested by users. So I agree with @qubvel on that both should be created on the same device as pixel_values.device, which can be "cuda" or "cpu" depending on input kwargs