I'm not sure why the code was written the way it was, but it seems we
can do better by letting GHC directly reuse `ReaderT`'s instances as
well as relaxing the constraints from

    instance Monad m       => Functor     (InputT m)
    instance Monad m       => Applicative (InputT m)
    instance Monad m       => Monad       (InputT m)

to

    instance Functor     m => Functor     (InputT m)
    instance Applicative m => Applicative (InputT m)
    instance Monad       m => Monad       (InputT m)

This looks the same on GHC 7.8 (pre-AMP) as well as GHC 7.10 (post-AMP)

Haddock displays those instances (which are in actually hidden to the user).
This makes it seem like InputT implements MonadState/Reader from the mtl
package.

The main benefit is to move all encoding-related functions
to the Posix module.

NOTE: This patch breaks the build on Windows, for now.

Instead, provide explicit get/put/modifyHistory actions.
This makes it easier to use InputT in a stack of monad transformers.

The former is already used by several other APIs; and the terminology "mode" is
already used elsewhere in Haskeline itself.

On Windows, preferTermMode is equivalent to defaultMode.  That issue will be fixed
in an upcoming patch.

This makes the terminal-style versus file-style behaviors explicit.
Temporarily breaks the Windows backend.

  Summary of changes:
   - Removed RankNTypes; added Rank2Types, UndecidableInstances
   - Revamped Command interface to be simpler and more monadic
   - Refactored key/command processing into new module RunCommand
   - Begin infrastructure for emacs/vi modes to have their own state.
   - Make Monad boilerplate a little simpler
   - TODO: The Windows build is broken.

Ensure that we move past the line after finishing, even if an exception occurs that we didn't control (for example, darcs or ghci's SIGINT handlers).

Add new MonadException module and class; now we use mtl's StateT and ReaderT, and it's easier for other users of those transformers to integrate with this library.