The pre-subsuption and pre-matching functions should NEVER make bogus
bindings of type variables, although they are free to bale out and make
too few bindings.

I hadn't been thiking carefully enough about this, and there were two
separate bugs.  

- Firstly, in pre-subsumption we must ignore the 'theta'
  part of any overloaded type.  

- Second, in pre-matching, we must return the empty subustition 
  on a mis-match, rather than returning the substitution so far.

This bug showed up when compiling Data.Generics.Schemes.hs, and is
imortalised in test tc206