Change example in "Copying data is not always bad" section of Performance Tips

[LilithHafner]

The previous example shows only a minor speedup for copying, and further, I could not reproduce that speedup at all when copying the example into a 1.7 or a 1.9 REPL. I couldn't find any example where copying the data sped up operation unless it was accessed repeatedly. Consequently, I changed the example to one with repeated access, and changed the prose to reflect the significance of repeated access. This makes sense to me because it seems difficult to recoup the cost of irregular access during a copy by avoiding a single irregular access during operation.

Repeated multiplication is silly because it makes more sense to raise A to a power first, so I added a ReLU activation function and made the example into a full neural network. It's the simplest sensible example I could come up with.