Tweak GoogLeNet and Inception family to match the torchvision implementations

[theabhirath]

These changes will make it easier to port over pretrained weights for the models from PyTorch.

• [ ] Right now, GoogLeNet matches the implementation in the paper, which does not use batch normalisation. However, torchvision uses batch normalisation layers after the convolution layers in both the inception block and the stem. It also uses bias = false for the convolution layer and eps = 0.001 for the batch normalisation layer. We could decide if similar to VGG, we can have a toggle for the batch normalisation in the model.
• [x] InceptionV3, InceptionV4 and InceptionResNetv2 have the right placement of normalisation layers but similar to GoogLeNet, use bias = false for the convolution layer and eps = 0.001 for the batch normalisation layer. (should be closed by #198)

[darsnack]

A toggle would be a good idea so that the paper variant is still easily available.

[pri1311]

Hi, this issue seems like a good place to start contributing.

function convolution(kernel_size, inplanes, outplanes, batchNorm; kwargs...)
    if batchNorm
        return basic_conv_bn(kernel_size, inplanes, outplanes; kwargs...)
    else 
        return Conv(kernel_size, inplanes => outplanes; kwargs...)
    end
end

would implementing such a function and the calling it like this convolution((7, 7), inchannels, 64, batchNorm; stride = 2, pad = 3, bias = bias) be a good way to refactor code for the particular issue? Or do you have something else in mind?

[darsnack]

We already have support for turning off the batch norm here. The remaining task on this issue it to update the code for GoogLeNet to use that functionality.