Deprecate older VGG API

[theabhirath]

Fix #269. Also adds the pre-trained API to the docstring of VGG. And a small rewrite of the VGG block, which was written pre-conv_norm's current more powerful form.

Right now the depwarn isn't printed by default, should we force it to be true? Or is it reasonable to assume that people will spot this and not upgrade if they don't want to/migrate to the newer API if they want to?

[github-actions]

Once the documentation build has completed, you can preview any updated documentation at this URL: https://fluxml.ai/Metalhead.jl/previews/PR270/

[theabhirath]

P.S. there has to be a way to limit the tests run when we open a PR, because running ResNet/MobileNet tests for a VGG PR seems hugely wasteful. Is there a hook to GH Actions that can let us specify which tests to run via a comment or something?

[darsnack]

I disagree—passing tests always seem wasteful 😁. IMO tests that don't always run are as bad as no tests at all.

[theabhirath]

I disagree—passing tests always seem wasteful 😁. IMO tests that don't always run are as bad as no tests at all.

That makes sense 😂 I was thinking more from the GH actions perspective of how many free minutes we get but I think the individual test sets are executing pretty quickly now so it should be okay

[darsnack]

Public repos have unlimited minutes, so that's not an issue. I guess the only thing is that a single org shares runners across all repos, and there are a fixed amount of runners you can use in parallel. But I haven't noticed this to be an issue; it's not like we are waiting significantly longer to merge PRs.

Shorter turnaround time is always nicer, but I say we target the actual source of latency for now.

[theabhirath]

Public repos have unlimited minutes, so that's not an issue.

Oh, for some reason I thought this was limited 😅 That's perfect then, waiting a bit for a PR to merge is not a problem at all.