[AutoDiff] Add loop differentiation negative testcases.

[dan-zheng]

Loop differentiation produces incorrect results when the reduction accumulation
variable is not initialized with an active parameter, e.g. when using
var result = 1 instead of var result = x.

func for_loop_nonactive_initial_value(_ x: Float) -> Float {
  var result: Float = 1
  for _ in 0..<2 {
    result = result * x
  }
  return result
}
print(valueWithGradient(at: 3, in: for_loop_nonactive_initial_value))
//   Actual: (value: 9.0, gradient: 3.0)
// Expected: (value: 9.0, gradient: 6.0)

TF-933 tracks this issue. This patch add negative testcases (currently failing).

---

Workarounds are to reformulate using var result = x or by using
Array.differentiableReduce:

// Workaround 1: use `var result = x` instead of `var result = 1`.
func for_loop_active_initial_value(_ x: Float) -> Float {
  var result = x
  for _ in 0..<1 {
    result = result * x
  }
  return result
}
print(valueWithGradient(at: 3, in: for_loop_active_initial_value))
// (value: 9.0, gradient: 6.0)

// Workaround 2: use `Array.differentiableReduce`.
func power(_ x: Double, _ exponent: Int) -> Double {
  // Array allocation isn't efficient, but `Array.differentiableReduce` has correct derivatives.
  let array = Array(repeating: x, count: exponent)
  return array.differentiableReduce(1, *)
}
print(valueWithGradient(at: 3, in: { x in power(x, 2) }))
// (value: 9.0, gradient: 6.0)

[dan-zheng]

@swift-ci Please test tensorflow