Tighten performance numbers for methods mentioned in our paper about bytestring-fiat

[jwiegley]

I've analyzed performance, and we're only weak in two places unpack and uncons, with the former having the greatest divergence.

The reason for this is that ByteString plays some really devious tricks, using a function it calls accursedUnutterablePerformIO. The comment for it is quite humorous: https://github.com/haskell/bytestring/blob/master/Data/ByteString/Internal.hs#L566

Although we could do the same as they did to match performance, it would require axiomatizing some more things, about whose semantics we know even less, so I think it's worth the 15% performance penalty for the time being.