Consider (Trac #15215)
  data T a = MkT ...
  data S a = ...T...MkT....

If there is an error in the definition of 'T' we add a
"fake type constructor" to the type environment, so that we
can continue to typecheck 'S'.  But we /were not/ adding
a fake anything for 'MkT' and so there was an internal
error when we met 'MkT' in the body of 'S'.

The fix is to add fake tycons for all the 'implicits' of 'T'.
This is done by mk_fake_tc in TcTyClsDecls.checkValidTyCl,
which now returns a /list/ of TyCons rather than just one.

On the way I did some refactoring:

* Rename TcTyDecls.tcAddImplicits to tcAddTyConsToGblEnv
  and make it /include/ the TyCons themeselves as well
  as their implicits

* Some incidental refactoring about tcRecSelBinds. The main
  thing is that I've avoided creating a HsValBinds that we
  immediately decompose.  That meant moving some deck chairs
  around.

NB: The new error message for the regression test T15215
has the opaque error "Illegal constraint in a type:", flagged
in Trac #14845.  But that's the fault of the latter ticket.
The fix here not to blame.