[AutoDiff] Add differentiability_witness_function instruction.

[dan-zheng]

differentiability_witness_function looks up a differentiability witness
function (JVP, VJP, or transpose) for a referenced function using SIL
differentiability witnesses.

sil [ossa] @generic : $@convention(thin) <T> (@in_guaranteed T, Float) -> @out T

sil @test_derivative_witnesses : $... {
  %vjp = differentiability_witness_function [vjp] [parameters 0 1] [results 0] <T where T: Differentiable, T == T.TangentVector> @generic : $@convention(thin) <T> (@in_guaranteed T, Float) -> @out T
}

The differentiability witness function kind identifies the witness function to
look up: [jvp], [vjp], or [transpose].

The remaining components identify the SIL differentiability witness:

• Original function name.
• Parameter indices.
• Result indices.
• Witness generic signature (optional). When parsing SIL, the parsed witness
  generic signature is combined with the original function's generic signature
  to form the full witness generic signature.

---

Known caveats:

• It may be problematic for SIL instructions like DifferentiabilityWitnessFunctionInst to store a SILFunction * pointer to the original function because SILFunction has special reference counting logic: see SILInstruction::dropAllReferences.
• It may be desirable to change the design of DifferentiabilityWitnessFunctionInst to directly store a SILDifferentiabilityWitness * pointer, or store the original function name and type. Deferring exploration until changes become necessary or the right design becomes evident.

---

The TF-866 master issue tracks all retroactive derivative registration tasks.

[dan-zheng]

@swift-ci Please test tensorflow