Do Standard 2D Convolution Kernels still require padding in specific cases?

[Zhao-Dongyu]

I believe that the text in the 3.2.1 Standard 2D Convolution Kernels section is very misleading.

I found this issue in the floating-point model. When the number of input channels is not a multiple of 2, padding was required before v2.3, but padding will cause errors in v2.3 and v3.0.

Could you please check this? Is it because the documentation has not been updated in time?