Add ONNX export support for PatchTST

[xenova]

What does this PR do?

This PR adds ONNX export support for PatchTST models, as well as fixes other models that contain the torch.unfold operator, which is not supported by torch.onnx.export. This is done by applying a set of default patching specs.

Fixes huggingface/transformers.js#1047

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@michaelbenayoun, @echarlaix

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[xenova]

I ran into some issues with my naive attempt (only seems to work for 4D). Here's a better version I found:

Suggested change

	def onnx_compatible_unfold(self, dimension, size, step):
	num_patches = (self.size(dimension) - size) // step + 1
	return torch.stack(
	[self[:, i : i + size, :] for i in range(0, num_patches * step, step)],
	dim=1,
	).transpose(3, 2)
	def onnx_compatible_unfold(input_tensor, dimension, size, step):
	"""
	Custom implementation of torch.unfold without using torch.unfold.
	Args:
	input_tensor (torch.Tensor): The input tensor.
	dimension (int): The dimension to unfold.
	size (int): The size of each slice.
	step (int): The step size between slices.
	Returns:
	torch.Tensor: The unfolded tensor.
	"""
	# Compute the shape of the unfolded output
	input_size = input_tensor.size(dimension)
	num_slices = (input_size - size) // step + 1
	# Permute dimension to the end for easier indexing
	input_tensor = input_tensor.transpose(dimension, -1)
	# Extract slices
	slices = []
	for i in range(num_slices):
	start = i * step
	end = start + size
	slices.append(input_tensor[..., start:end])
	# Stack slices and permute dimensions back
	result = torch.stack(slices, dim=-2).transpose(dimension, -2)
	return result