TODO

 - Better docs for why plans are useful.
 - Mention that dct2 is "the" DCT.
 - Figure out why TimeFFTW.hs and time-fftw.c don't match up in timings.
 - Better docs for c2r/r2c in all modules.
   Specifically, note that it's orthog/invert only when first and last are real (I
   think?)
 - Multidimensional transforms.
 - Bench and optimize the normalized transforms.
 - Explicit versions in build-depends.
 - Maybe better to have normalizeInput/output :: Vector -> Vector?
 - Rest of the orthonormal transforms.
 - Tests that I'm using all of vector's unsafe ops correctly.
 - Tests for Unnormalized, that <Fu,v>==<u,F^*v>.
 - Can we get more accuracy wrt normalizations?
 - Add a note that this doesn't return [1..5] (b/c output is even)
        run dftC2R $ run dftR2C $ fromList [1..5]
   and must be replaced by:
        execute (plan dftC2R 5) $ run dftR2C $ fromList [1..5]
 - Create performance benchmarks using criterion.
 - Get the inlining to work properly, both when using "run" and when using
   "plan" plus several executes.
 - Size-zero input vectors?
 - Is -fexcess-precision necessary?
 - Better timings:
   - Bench (and optimize) the normalizations.
   - Put Plans in a list and apply them separately (to prevent unboxing), and
     see how much of a performance hit it is.
 - QuickCheck properties matching each of the transforms against their (slow)
   mathematical definition.