* More thorough tests of the delta generation.  I've only tried it on
   a few things, so it's probably full of bugs.

 * It won't work with files bigger than 2Gb, becuase:
    * COPY operations for more than 2Gb need to be split up into multiple
      commands.
    * The COPY opcode 255 (which has a 64 bit offset value) isn't
      implemented yet in either delta generation or application.

 * There are a few places where the algorithm doesn't choose the optimal
   delta but where it wouldn't be too hard to help it along.  The ones
   I've spotted would only save a tiny amount of space though.

 * Document the algorithm, because that will help me know how to improve it.

 * Do something about file buffering.  It could probably make delta
   generation massively faster if we keep buffers around longer.

 * Add support for passing references to scalars in instead of file
   handles.  That should be much more efficient than using IO::Scalar.
   Then remove the note suggesting this from the POD.

 * It probably leaks memory everywhere.  Fix it up and valgrind it.

 * Proper testing of some of the inner bits in test.c

 * require IO::Scalar? or just recommend?

 * Make sure that 2.gdiff is generated during release process automatically.