[Sema] Differentiable conformance derivation for class types.

[dan-zheng]

Differentiable derived conformances now supports class types.
Synthesis works just like for struct types.

Class differentiation support requires further differentiation transform changes.

Resolves TF-630.

Next steps:

• TF-631: Enable differentiation of class methods (handle ref_element_addr in differentiation transform).
• TF-633: Update tensorflow-swift-apis using Differentiable-conforming class types.

---

class Example<T : Differentiable> : Differentiable {
  var x, y: T
  @noDerivative var flag: Bool = true

  // Unlike structs, classes do not get synthesized memberwise initializers.
  // User must define initializer or provide default values for stored properties.
  init(x: T, y: T) {
    self.x = x
    self.y = y
  }

  // Compiler synthesizes (omitting `AllDifferentiableVariables` code):
  //
  // struct TangentVector : Differentiable, AdditiveArithmetic {
  //   var x: T.TangentVector
  //   var y: T.TangentVector
  //   init(x: T.TangentVector, y: T.TangentVector)
  //   typealias TangentVector = Example<T>.TangentVector
  //   final static var zero: Example<T>.TangentVector { get }
  //   static func + (lhs: Example<T>.TangentVector, rhs: Example<T>.TangentVector) -> Example<T>.TangentVector
  //   static func - (lhs: Example<T>.TangentVector, rhs: Example<T>.TangentVector) -> Example<T>.TangentVector
  //   @_implements(Equatable, ==(_:_:)) static func __derived_struct_equals(_ a: Example<T>.TangentVector, _ b: Example<T>.TangentVector) -> Bool
  // }
  //
  // func move(along direction: Example<T>.TangentVector) {
  //   x.move(along: direction.x)
  //   y.move(along: direction.y)
  // }
}

[dan-zheng]

These tests are not the best, I directly copied test/Sema/struct_differentiable.swift.
Todo: clean up and tighten tests. Fuzzing would be cool.

[dan-zheng]

@swift-ci Please test tensorflow

[rxwei]

Could you add a test that guarantees that TangentVector will never be derived to be Self for classes, even if Self conforms to AdditiveArithmetic?

[dan-zheng]

Could you add a test that guarantees that TangentVector will never be derived to be Self for classes, even if Self conforms to AdditiveArithmetic?

I actually didn't implement that logic, will do so now.
Done in 008789b.

[dan-zheng]

@swift-ci Please test tensorflow