diff: Fix bug in modechecking of non-local lambda vars

[Andrew Bromage]

G'day all.

Fergus Henderson wrote:

> > +		% XXX This test is too conservative.
> > +		%
> > +		%     We should allow non-local variables to be non-ground
> > +		%     if they are dead after this unification.  In addition,
> 
> Why only if they are dead after this unification?
> Why not always?

I suppose it depends what you consider to be a violation of variable
locking.  This requires some thought.

Is it a violation of variable locking to:

	- Bind a free variable and then backtrack, so the var was
	  never actually bound?

	- Do almost anything with an `any' variable?

	- Alias a free variable to an `aliased_many' state, which
	  requires some handle on indirections (e.g. the circular
	  list representation)?

Operationally, if a variable is locked by some goal, you shouldn't be
allowed to change that variable's low-level representation of a locked
var (or, at least, it should be put back as it was at the end of that
goal, which suggests that there might be a difference on whether or not
we're in a trailing grade).  How does this map onto the high-level insts
in a way that is reasonable for different underlying implementations?

> Otherwise, that looks good.  But can I please see
> another diff before you commit that?  Thanks.

Will do.  Just after we've sorted out the above bits.  All your other
suggestions have been implemented.

Thanks, Fergus.

Cheers,
Andrew Bromage