Implementation of Inceptionv4, InceptionResNetv2 and Xception

[theabhirath]

This PR is my first official contribution as a GSoC student 😁. It does a couple of things:

1. It adds new models (Inceptionv4 and InceptionResNetv2). Internal docs for the layers of these models are left out because they are repetitive and don't really explain much - but if that's a concern I can always fill them in.
2. It deprecates Inception3 in favour of Inceptionv3 since the vX notation is being used for models in other libraries (as well as in Metalhead for MobileNetv3) and so it makes sense to be explicit about it.

Other things that I happened to fill in because I found them:

1. cat_channels is now type-stable as it uses Val(3) for dims - thanks to @ToucheSir pointing that out in some conversations
2. Catches some formatting issues with Hotfix for ViT on GPU #169

Notes:

1. TTFG for the Inception models is off the charts. Particularly, InceptionResNetv2 is absolutely insane, sometimes taking minutes. Subsequent gradients are quite slow as well (in comparison, ViTs, which are heavier models, seem to be faster) and also take a lot of memory. This might be helped by using Chain(::Vector) but I haven't used that yet because I'm not sure if it is causing stability issues while training.
2. Closes Implementation of Inceptionv4 #131 (supersedes). Should also close Implementation of Inception v4 and Inception resnet v2 #129 since that hasn't seen activity for quite some time.

Edit: added the Xception model as well

[darsnack]

Looks nearly ready right off the bat, nice job! Just a couple initial comments.

[theabhirath]

CI times are through the roof...we might have to disable gradtests again (or maybe enable it only for select models)

[theabhirath]

CI times are through the roof...we might have to disable gradtests again (or maybe enable it only for select models)

...how on earth are the second CI times so much lesser than the first one?

[ToucheSir]

AIUI GHA runners are containerized but not otherwise isolated, so times can vary drastically depending on what else is running on the host.

[theabhirath]

Xception seems to have an even worse gradient benchmark:

julia> x = rand(Float32, 299, 299, 3, 2);

julia> model = Xception();

julia> @benchmark Zygote.gradient(p -> sum($model(p)), $x)
BenchmarkTools.Trial: 1 sample with 1 evaluation.
 Single result which took 9.255 s (78.29% GC) to evaluate,
 with a memory estimate of 10.65 GiB, over 1553316 allocations.

[theabhirath]

This should be ready now, although I'm not sure what exactly I am to make of the gradient times. I suppose they shouldn't block this PR, but is there something that could be a possible contributor that I can change?

[theabhirath]

The memory problems return 😬

[darsnack]

Not sure there's much we can do beyond what's already been tried in Metalhead. We just need to do something about applychain in Flux/Zygote.

[darsnack]

Only some minor doc stuff

[darsnack]

Is there a use case of (true, false) or (false, true)? If not, I was thinking it is kinda silly to use conv_bn with the keyword use_bn = false cause that's just Conv. It makes sense for consistency with depthwise_sep_conv_bn. But it might be cleaner to just introduce depthwise_sep_conv then instead of more keywords in multiple places?

[theabhirath]

I can't say I see something yet but I think this can be kept for some time...with all the churn over the next two months it might be handy. If not, I'll simplify it later

[theabhirath]

Ah sorry, I keep missing these minor doc tweaks 🤦🏽 I'll be a bit more thorough next time

[darsnack]

No problem! Thanks and nice job!