Fix initial value handling with flatten transducer (#34369)

There was a bug in initial value handling of `FlatteningRF` and the
following example failed:

    @test mapfoldl(
        x -> (x, x),
        ((a, b), (c, d)) -> (min(a, c), max(b, d)),
        Iterators.flatten((1:2, 3:4)),
    ) == (1, 4)

This is because `BottomRF(op.rf)` was called inside `FlatteningRF`
where `op.rf` is already a "non-bottom" reducing function; here it's a
`MappingRF`.  As `BottomRF(rf)` forwards anything on the second
argument on the first invocation as the first argument (accumulator)
of the next calls, we need to make sure that this value is processed
through `MappingRF` in the above example.  However, if we do
`BottomRF(op.rf)` where `op.rf` is a `MappingRF`, this `BottomRF`
bypasses any processing that has to happen in `op.rf`.

(cherry picked from commit 0ee3264)

Author: tkf

base/reduce.jl

@@ -112,7 +112,7 @@ struct FlatteningRF{T}
 end
 
 @inline function (op::FlatteningRF)(acc, x)
-    op′, itr′ = _xfadjoint(BottomRF(op.rf), x)
+    op′, itr′ = _xfadjoint(op.rf, x)
     return _foldl_impl(op′, acc, itr′)
 end
 

test/reduce.jl

@@ -535,3 +535,11 @@ x = [j^2 for j in i]
 i = Base.Slice(0:0)
 x = [j+7 for j in i]
 @test sum(x) == 7
+
+@testset "initial value handling with flatten" begin
+    @test mapfoldl(
+        x -> (x, x),
+        ((a, b), (c, d)) -> (min(a, c), max(b, d)),
+        Iterators.flatten((1:2, 3:4)),
+    ) == (1, 4)
+end