[AutoDiff] Make force-unwrapping differentiable.

[rxwei]

Optional force-unwrapping is a mathematically transposable operation. This patch makes force-unwrapping differentiable by applying its transpose on adjoint buffers.

func bla<T: Differentiable & FloatingPoint>(_ t: T) -> (T, Float) where T == T.TangentVector {
    gradient(at: t, Float(1)) { (x, y) in (x as! Float) * y }
}
print(bla(Float(2))) // (1, 2)

Resolves TF-455.

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow Linux