[OnnxToTorch] Fix Resize op when ONNX exports dynamic spatial dims as 0

[HalfBloodPrince010]

ONNX encodes dynamic spatial dimensions in Resize as 0. This pass just passes the spatial dimensions as is for next steps. For interpolate/resize op, [0,0] spatial dimension is not valid proposed size. This change replaces such 0 values with the corresponding runtime dimension from the input tensor, ensuring correct shape propagation in Torch-MLIR and preventing invalid 0-sized dimensions.

This fixes the issue: iree-org/iree#19501

Example for ONNX Model Export

>>> import torch
>>> import torch.nn as nn
>>> import torch.onnx
>>>
>>> class InterpolateModel(nn.Module):
...     def forward(self, x):
...         return nn.functional.interpolate(x, scale_factor=0.5, mode='bicubic', align_corners=False)
...
>>> model = InterpolateModel()
>>> dummy_input = torch.randn(11, 3, 32, 54)
>>>
>>> torch.onnx.export(
...     model,
...     dummy_input,
...     "interpolate_dynamic_cubic.onnx",
...     opset_version=11,
...     input_names=['pixel_values'],
...     output_names=['output'],
...     dynamic_axes={
...         'pixel_values': {0: 'batch', 2: 'height', 3: 'width'},
...         'output': {0: 'batch', 2: 'height', 3: 'width'},
...     }
... )

After exporting the model
>>>
>>> print("Exported interpolate_dynamic_cubic.onnx")
Exported interpolate_dynamic_cubic.onnx
>>> dynamic_input_model = onnx.load("interpolate_dynamic_cubic.onnx")
>>> for input_tensor in dynamic_input_model.graph.input:
...     shape = [dim.dim_value for dim in input_tensor.type.tensor_type.shape.dim]
...     print(f"{input_tensor.name}: {shape}")
...
pixel_values: [0, 3, 0, 0]

In IREE, one of the test using Hugging Model is failing because the interpolate/resize
dimensions are [0, 0]

1. Input (pixel_values)
2. Get Shape
3. Gather indices 2 and 3 (which is spatial dimension height and width)
4. Divide both by 16 (patch size)
5. Concat them.

But input pixrl_values are all [0, 0 , 0, 0]

import onnx
model = onnx.load("model.onnx")

>>> for input_tensor in model.graph.input:
...     shape = [dim.dim_value for dim in input_tensor.type.tensor_type.shape.dim]
...     print(f"{input_tensor.name}: {shape}")
...
pixel_values: [0, 0, 0, 0]