Remove @differentiable attribute from Layer.inferring(from:). (#575)

Necessary for TF-835: lowering `@derivative` attribute directly to SIL
differentiability witnesses instead of creating implicit `@differentiable`
attributes.

`@differentiable` + `@derivative` attributes for the same original declaration
with the same wrt parameters will be diagnosed.

To relax this limitation, derivative generic signature mangling (TF-680) is
necessary to avoid name collisions for SIL derivative functions with the same
parameter indices but different derivative generic signatures.

Author: dan-zheng

Sources/TensorFlow/Layer.swift

@@ -85,12 +85,12 @@ public extension Layer {
     ///
     /// - Parameter input: The input to the layer.
     /// - Returns: The inference output.
-    @differentiable
     func inferring(from input: Input) -> Output {
         withLearningPhase(LearningPhase.inference) { self(input) }
     }
 
-    // TODO(rxwei): Remove this custom VJP once differentiation supports currying.
+    // TODO(TF-433, SR-11882): Remove this custom derivative when
+    // differentiation supports `rethrows` functions and currying.
     @derivative(of: inferring(from:))
     @usableFromInline
     internal func _vjpInferring(from input: Input)