[AutoDiff] Add #chainableGradient differential operator for seedable gradient

[rxwei]

#chainableGradient takes a function and returns a function, which takes original parameters and a backpropagated adjoint, and computes vector-Jacobian products.

func f(_ x: Float) -> Float { ... }
let y = f(x)
let dz_dy = ...
let dz_dx = #chainableGradient(f)(x, dz_dy)

#chainableGradient corresponds to the gradient [seedable] instruction in SIL.

Sema and libSyntax are to be done separately. This is a random patch to unblock the progress towards a way to express seedable derivatives in Swift source code.

[rxwei]

@swift-ci please test tensorflow

[dan-zheng]

Sweet!

[dan-zheng]

I'm happy to tackle Sema and Syntax for #chainableGradient if you'd like to work on other parts of autodiff. Let me know!

[rxwei]

@dan-zheng Thx. Don't worry about it today. Let's try to get more deabstraction bugs fixed.

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow