Remove MonadFail (ST s) instance

[Bertram Felgenhauer]

> > It's certainly not in the spirit of the MonadFail proposal. Arguably,
> > *all* implementations of `fail` in terms of `error*` are wrong, even
> > the one for IO. (I suspect dropping the instance for IO would cause a
> > lot of pain though...)
> 
> Isn't fail in IO a synonym for (ioError . userError) ?

Oh right. I guess the IO instance is okay then.

Thanks for the correction,

Bertram