fix performance of ProductIterator for non-concrete iterators (#59711)

This uses inferred element types to make the iterator state type-stable,
enabling inlining, which can enable the runtime to see the real types.
This is essentially the transform that we want inference to infer and
codegen to implement, except made explicit so that we can rely on it
actually happening.

Fix #56607

```
julia> @Btime f()
  574.500 ns (2 allocations: 272 bytes) # non-const op
  253.878 ns (2 allocations: 272 bytes) # const op
```

(cherry picked from commit 498d982)

Author: vtjnash

base/iterators.jl

@@ -16,7 +16,7 @@ using .Base:
     (:), |, +, -, *, !==, !, ==, !=, <=, <, >, >=, =>, missing,
     any, _counttuple, eachindex, ntuple, zero, prod, reduce, in, firstindex, lastindex,
     tail, fieldtypes, min, max, minimum, zero, oneunit, promote, promote_shape, LazyString,
-    afoldl
+    afoldl, @default_eltype
 using Core
 using Core: @doc
 
@@ -1140,6 +1140,8 @@ end
     next === nothing && return nothing
     restnext = _piterate(rest...)
     restnext === nothing && return nothing
+    VS = @default_eltype(iter1)
+    next = Pair{VS, typeof(next[2])}(next[1], next[2])
     return (next, restnext...)
 end
 @inline function iterate(P::ProductIterator)
@@ -1152,15 +1154,16 @@ end
 @inline _piterate1(::Tuple{}, ::Tuple{}) = nothing
 @inline function _piterate1(iters, states)
     iter1 = first(iters)
-    next = iterate(iter1, first(states)[2])
-    restnext = tail(states)
+    state1, restnext... = states
+    next = iterate(iter1, state1[2])
     if next === nothing
         isdone(iter1) === true && return nothing
         restnext = _piterate1(tail(iters), restnext)
         restnext === nothing && return nothing
         next = iterate(iter1)
         next === nothing && return nothing
     end
+    next = Pair{fieldtype(typeof(state1), 1), typeof(next[2])}(next[1], next[2])
     return (next, restnext...)
 end
 @inline function iterate(P::ProductIterator, states)