fix bug in merge_kernel for structural_reparam

[liyiersan]

When applying model.structural_reparam(), there is a need to add padding so that the output size can match, see #48.
Besides, I find that the kernel_size, stride, padding, and dilation may be tuples.

def get_conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias):
      # other codes
      if has_large_impl and in_channels == out_channels and out_channels == groups and use_large_impl and stride == 1 and padding == kernel_size // 2 and dilation == 1:
      # kernel_size, stride, padding, and dilation may be tuples, the condition may be False!
      # other codes

Therefore, we need to convert the tuple to int to make DepthWiseConv2dImplicitGEMM work, or it would be replaced with nn.Conv2d.