2-D Convolution Forward Pass

Implement the forward pass of a 2-D convolution layer (cross-correlation, as frameworks use).

conv2d(x, w, b, stride=1, padding=0):

• x: input of shape (C_in, H, W)
• w: kernels of shape (C_out, C_in, KH, KW)
• b: bias of shape (C_out,)
• Zero-pad the height and width by padding on each side, then slide each kernel with the given stride.

Output shape (C_out, H_out, W_out) with $H_{out} = \lfloor (H + 2p - KH)/s \rfloor + 1$ (and the same for width). Loop over output positions if you like, but vectorise over channels.