Fix input processing for convolution

[Cecilwang]

Thanks for the previous support for convolution layers. In this PR, I’ve added fixes for two specific cases:

• Conv1D and unfold parameter requirement:
  The unfold operation expects 2D parameters, so for Conv1D layers, both the input tensor and parameters need to be expanded by one dimension to match the expected shape.

• Grouped convolutions:
  Fixed the input tensor shape for convolutions that use the groups parameter to ensure the input is correctly formatted and grouped during computation.

ref:
https://pytorch.org/docs/stable/generated/torch.nn.Unfold.html
https://discuss.pytorch.org/t/conv1d-implementation-using-torch-nn-functional-unfold/109643/3

[Qubitium]

@Cecilwang Do you know of a opensource model currently on HF that would trigger this code? I want want to add test_conv1d.py tests to CI so we can check for regressions in the future. I know you are currently applying to a private model but was wondering if you know a non-private model that we can do a simple quant/inference on?

Otherwise, this PR is great and ready to go!

[Cecilwang]

state-spaces/mamba-130m-hf is a MambaForCausalLM whose slow_forward function might trigger this PR's logical. You can try it without installing causal_conv1d.

https://github.com/huggingface/transformers/blob/main/src/transformers/models/mamba/modeling_mamba.py#L54-L61
https://github.com/huggingface/transformers/blob/main/src/transformers/models/mamba/modeling_mamba.py#L326-L328
https://github.com/huggingface/transformers/blob/fee1190601b5d04ec6d3f7f58fd22788d7f3236d/src/transformers/models/mamba/modeling_mamba.py#L274