Fix ParameterlessLayer conformances. (#620)

Associated type inference behavior was changed in
swiftlang/swift#32578:
derived conformances are now attempted before associated type inference.

This broke `ParameterlessLayer`, which relied on a
`TangentVector == EmptyTangentVector` same-type constraint to set a
default `TangentVector` type witness for conforming types.

Add explicit `TangentVector` type witnesses to
`ParameterlessLayer`-conforming types to fix this regression.

This workaround is forward- and backward-compatible, but makes code more
verbose.

Author: dan-zheng

FastStyleTransfer/Layers/Helpers.swift

@@ -2,6 +2,8 @@ import TensorFlow
 
 /// A 2-D layer applying padding with reflection over a mini-batch.
 public struct ReflectionPad2D<Scalar: TensorFlowFloatingPoint>: ParameterlessLayer {
+    public typealias TangentVector = EmptyTangentVector
+
     /// The padding values along the spatial dimensions.
     @noDerivative public let padding: ((Int, Int), (Int, Int))
 
@@ -39,6 +41,8 @@ public struct ReflectionPad2D<Scalar: TensorFlowFloatingPoint>: ParameterlessLay
 
 /// A layer applying `relu` activation function.
 public struct ReLU<Scalar: TensorFlowFloatingPoint>: ParameterlessLayer {
+    public typealias TangentVector = EmptyTangentVector
+
     /// Returns the output obtained from applying the layer to the given input.
     ///
     /// - Parameter input: The input to the layer.

Models/ImageClassification/ShuffleNetV2.swift

@@ -19,6 +19,8 @@ import TensorFlow
 // Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, Jian Sun
 
 public struct ChannelShuffle: ParameterlessLayer {
+    public typealias TangentVector = EmptyTangentVector
+
     @noDerivative public var groups: Int
 
     public init(groups: Int = 2) {

Models/Text/GPT2/TransformerLM.swift

@@ -130,6 +130,8 @@ func _vjpCausallyMasked(_ dotProducts: Tensor<Float>, enable: Bool)
 }
 
 struct Attention: ParameterlessLayer {
+    typealias TangentVector = EmptyTangentVector
+
     @noDerivative var dropout: Dropout<Float>
     @noDerivative var scale: Tensor<Float>
     @noDerivative var causal: Bool

pix2pix/Layers.swift

@@ -15,6 +15,8 @@ import TensorFlow
 import Foundation
 
 public struct Identity: ParameterlessLayer {
+    public typealias TangentVector = EmptyTangentVector
+
     @differentiable
     public func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> {
         input