Over specialization on function argument

[yuyichao]

Ref #15177 (comment)
On my laptop, the following code takes < 0.02s on 0.4 and > 40s on master.

function f(m)
    for i in names(m)
        v = m.(i)
        isa(v, Function) && string(v)
    end
end
@time f(Base)

According to the profiling result, it spends most of the time in inferencing and compiling string (or print_to_string) even though the result of the compilation isn't any useful since rerunning f(Base) takes roughly the same time on 0.4....

As Stefan pointed out, the heuristic of not specializing on what we don't call should solve this issue but somehow it is not triggerred in this case.

@jrevels for performance benchmark. In additional to the reduced case above, the first half of genstdlib.jl script before #15177 should be a good candidate too.

c.c. @JeffBezanson

[carnaval]

@JeffBezanson the problem here is MAX_UNSPECIALIZED_CONFLICTS, so even though the logic for "this argument is or is not being called in this function" is correct, it gives up on unspecializing the cache entry

[JeffBezanson]

Ok, let's bump it to 20.

My job is so easy :)