Add BangPatterns for an accumulator. Inline decode. Specialize take, drop.

When considering a simple benchmark [1]:

   let p = ((== "en") . U.take 2)
   print . length . filter p . U.lines =<< L.readFile . head =<< getArgs

It's apparent that we are underperforming (compare to the shell script
solution, this take twice as long).  The accumulator of splitAt wasn't
strict (used bang patterns).  Also, even with -O2, the construction /
destruction of the tuple for take & drop's lifting of splitAt wasn't
being optimized away.  Perhaps more investigation should be done re:
GHC, but the cheap solution is to just duplicate code for take and
drop, which is what this patch does.

Original: user    0m1.704s
Final: user    0m1.092s

[1] http://stackoverflow.com/questions/8172889/quickly-parse-large-utf-8-text-file-in-haskell