Add initial how_to_scale_op notebook

[DouglasOrr]

View notebook.

My thoughts for the blog post were to keep it pretty short, with some key figures & results, rather than repeating all of this there.

Feedback of any sort is very welcome - thanks in advance!

[thecharlieblake]

This looks great - thanks so much! A few things came to mind, though I appreciate the blog version might address some of it anyway:

1. A bit more context for the un-initiated might be nice up-front - i.e. a para on what u-µP is for someone who's never heard of it
2. I had never heard of hardtanh before this example - a sentence introducing it and why you chose it might help the reader
3. Also a sentence explaining that our lib has these lovely scale_fwd and scale_bwd fns
4. Generally I'm finding the maths a bit hard to follow - particularly the jump to E[y^2]. I think we might benefit from spelling this out quite a bit more explicitly, even if it is more verbose. I'm guessing most ML people won't have done much of this kind of analysis before