patch reduce_max and ConvertModel.forward retention of output activations

[sammlapp]

These were small patches that I needed to implement in order to convert an onnx model (link below) to PyTorch, as mentioned in #78
onnx model: https://huggingface.co/justinchuby/Perch-onnx/blob/main/perch_v2_no_dft.onnx

For what it's worth, I also had to modify the Pad and Reshape ConvertModel layers to avoid incorrect padding and reshaping behavior. I don't know if there is a generalizable fix to these issues.

# patches to fix ONNX → PyTorch conversion issues
class FixedONNXPad(nn.Module):  # patchs onnx2pytorch's pad.Pad
    def forward(self, input, pads=None, value=0):
        if pads is None:
            raise TypeError("pads must be provided")

        pads = pads.tolist() if torch.is_tensor(pads) else list(pads)

        ndim = input.ndim
        assert len(pads) == 2 * ndim, (pads, input.shape)

        before = pads[:ndim]
        after = pads[ndim:]

        # Convert ONNX → PyTorch pad order
        torch_pads = []
        for b, a in zip(reversed(before), reversed(after)):
            torch_pads.extend([b, a])

        return F.pad(input, torch_pads, value=value)


class FixedReshape(nn.Module):
    def forward(self, x, shape=None):
        B, H, Q, D = x.shape
        assert D % 4 == 0
        return x.view(B, H, Q, D // 4, 4)
    
onnx_model = onnx.load("./perch_v2_no_dft.onnx") 
pytorch_model = ConvertModel(onnx_model)
pytorch_model.Pad_pad_output_0 = FixedONNXPad()
pytorch_model._modules[
    "Reshape_jit(infer_fn)/MultiHeadClassifier/MultiHeadClassifier._call_model/heads_protopnet_logits/dot_general10_reshaped_0"
] = FixedReshape()