[release-0.5] use a counter for passed tests instead of storing the Result object

[maleadt]

Ref #20150, cherry-picked from b7c850e

Reason for backport: after converting CUDAdrv.jl to using testsets, some finalizers weren't being called anymore, but only on 0.5.

Repro:

using Base.Test

type Foo
    whatever    # prevent singleton object
    function Foo()
        obj = new(nothing)
        finalizer(obj, finalize)
        return obj
    end
end

finalized = false
function finalize(obj::Foo)
    global finalized
    finalized = true
end

@testset "dummy" begin
    dummy = Foo()
    @test_throws MethodError sin(dummy)
end

try throw(nothing) end
gc()
@assert finalized

This is semi important to me, because for CUDAdrv.jl (any other GPU packages too) finalizers are responsible for freeing GPU memory, and without this backport we can't test whether that works reliably.

cc @KristofferC

[KristofferC]

Don't think this should break anything so should be fine to backport.

[tkelman]

adding a field could cause issues if anyone is constructing the type directly

[maleadt]

PkgEval? Or I can do a GitHub search for DefaultTestSet and test matching packages after this change.

[maleadt]

Dug through the GitHub code search results and couldn't find anyone constructing a DefaultTestSet by specifying all fields.
@tkelman do you object to this change because it could be breaking?

[KristofferC]

This shouldn't influence any user facing API. In my opinion we have to allow things like this or else we can't backported anything, because any change in behaviour is potentially breaking for someone who relies on that behaviour.