Ensures that h1 gets inlined into its use sites in cases like:

"""
h1 k = k undefined undefined undefined
        undefined undefined undefined
        undefined undefined undefined
        undefined undefined undefined
        undefined undefined undefined
        undefined undefined undefined

a = h1 (\x _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ -> x)
b = h1 (\_ x _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ -> x)
"""

I've benchmarked this on nofib (albeit recompiling only the
benchmarks, not the library) and it hardly shifts the numbers - binary
size is up by 0.1% at most (average 0.0%) and the worst-case
allocation increase is 0.2% (best case -0.1%, 0.0% average).

If you also rebuild the libraries with this change, the only further
change is a +0.2% allocation increase in cacheprof. So this looks like
a pretty low-risk change that will considerably benefit certain
programs.