Change the transformer to norm_first by default

[angeloskath]

This is a backwards incompatible change as it changes the default to norm_first however I think this should have always been the default and hopefully if not too many people relied on it we caught it early.

If not we can cancel this PR.

[awni]

Yea at some point I reviewed a PR which added a bunch of stuff to Transformer including this. I think I said we should just leave it to post norm since it's way more common now (?).

But this gets us compatibility with PyTorch so I'm also on board switching. I don't think it should break too much. In most of our examples we reimplement the Transformer anyway 😅

[angeloskath]

Just to be clear. This breaks compatibility with PT and moves to the more common option. As you said the only thing that is used in places is the attention implementation and I think this is more of a guide/example at this point.

[awni]

Sorry I confused myself 😅 , I actually did the exact opposite of what I thought I did in this PR #167 , where it was switched to post norm by default.

I am fine with reverting that.

[awni]

@angeloskath feel free to merge this. I tend to prefer supporting the more common option (especially when it is much more common) over perfect compatibility with PyTorch