[AutoDiff] Fix stdlib @derivative(of: subscript) crash.

[dan-zheng]

Fix @derivative(of: subscript) crash for SIMD{n} types.
Strangely, the crash is not easily reproducible via standalone tests:

// Standalone reproducer: no crash.
 extension SIMD2 {
  @differentiable(
    where Scalar : EuclideanDifferentiable & BinaryFloatingPoint,
          Scalar.TangentVector : BinaryFloatingPoint)
  public subscript(index index: Int) -> Scalar {
    @_transparent get { fatalError() }
    @_transparent set { fatalError() }
  }
}
extension SIMD2
  where Scalar : EuclideanDifferentiable & BinaryFloatingPoint,
        Scalar.TangentVector : BinaryFloatingPoint {
  @usableFromInline
  @derivative(of: subscript(index:))
  internal func _vjpSubscript(index: Int)
  -> (value: Scalar, pullback: (Scalar.TangentVector) -> TangentVector) {
    return (self[index], { v in
      var zeros = Self.zero
      zeros[index] = Scalar(v)
      return zeros
    })
  }
}

Resolves TF-1090.

Exposes TF-1094: during .swiftinterface compilation: unserialized
@derivative function is not registered for serialized @differentiable
function.

---

Issue originally reported by @bartchr808 in 73d43cf.

[dan-zheng]

@swift-ci Please test tensorflow

[dan-zheng]

@swift-ci Please test tensorflow

[bartchr808]

Looks great! Looking back, making SIMD differentiable has brought to light several interesting bugs 😄

[dan-zheng]

@swift-ci Please test tensorflow