When specialising a top-level recursive group, if none of the binders
are exported then we can start specialising based on the later calls to
the functions.
This is instead of creating specialisations based on the RHS of the
bindings.
The main benefit of this is that only specialisations that will actually
be used are created. This saves quite a bit of memory when compiling
stream-fusion and ForceSpecConstr sort of code.

Nofib has an average allocation and runtime of -0.7%, maximum 2%.
There are a few with significant decreases in allocation (10 - 20%)
but, interestingly, those ones seem to have similar runtimes.
One of these does have a significantly reduced total elapsed time
though: -38%.

On average the nofib compilation times are the same, but they do vary
with s.d. of -4 to 4%.
I think this is acceptable because of the fairly major code blowup fixes
this has for fusion-style code.
(In one example, a SpecConstr was previously producing 122,000 term size,
now only produces 28,000 with the same object code)