We used to use the base package to do the detection. However, base has
    extra-lib-dirs: @ICONV_LIB_DIRS@
in base.buildinfo.in, which means that if ICONV_LIB_DIRS is set then
the package config will have
    library-dirs: [...]base/dist-install/build /path/for/iconv
and, as paths may be double quoted and contains spaces, this makes it
a little tricky to extra the path that we want to look in (the first
path).

We therefore now use the ghc-prim package, which has no such complication.

We were checking paths with
    if [ -f '"/path/to/Prelude"' ]
i.e. the " quotes were being quoted by the ' quotes. Now we just
use " quotes (which come from the ghc-pkg output).

One adjustment is to the core output of some tests (false positive).

The other adjustment is due to higher allocation when (now) unboxed
values are put into a polymorphic data structure. This is a problem
with polymorphic data structures in general (e.g. it happens in Java
too) and not about having unboxed fields per say.

In effect, the error context for naked variables now takes up
a "slot" in the context stack; but it is often empty.  So the
context stack becomes one shorter in those cases.  I don't think
this matters; indeed, it's aguably an improvement.  Anyway that's
why so many tests are affected.

Really, the existence or otherwise of the .hi file shouldn't affect validate
output, but it seems to.  Maybe 'make clean' doesn't clean enough?

In ay case peak memory usage is a fragile number because it depends on when
GC happens.

So I increased the range to 20% for now.