[AutoDiff] Support differentiation of functions with multiple results in SIL.

[dan-zheng]

Reverse-mode differentiation now supports apply instructions with multiple
active "semantic results" (formal results or inout parameters).

The "cannot differentiate through multiple results" non-differentiability error
is lifted.

Resolves TF-983.

---

Follow-up:

• TF-1288: relax @differentiable and @derivative type-checking to allow
  original declarations with multiple semantic results.
• Add forward-mode differentiation support for apply instructions with multiple
  active "semantic results".

---

Example:

import _Differentiation

// Test function returning a tuple of active results.
func tuple(_ x: Float, _ y: Float) -> (Float, Float) {
  return (x, y)
}
func multiply(_ x: Float, _ y: Float) -> Float {
  let z = tuple(x, y)
  // Note: both results (tuple elements) are active.
  return z.0 * z.1
}
print(valueWithGradient(at: 3, 4, in: multiply))
print(valueWithGradient(at: 5, 10, in: multiply))

// Test function with multiple `inout` parameters.
func swap<T>(_ x: inout T, _ y: inout T) {
  let tmp = x; x = y; y = tmp
}
func multiply_swap(_ x: Float, _ y: Float) -> Float {
  var tuple = (x, y)
  swap(&tuple.0, &tuple.1)
  return tuple.0 * tuple.1
}
print(valueWithGradient(at: 3, 4, in: multiply_swap))
print(valueWithGradient(at: 5, 10, in: multiply_swap))

Before:

$ swift test.swift
test.swift:25:40: error: function is not differentiable
print(valueWithGradient(at: 5, 10, in: multiply_swap))
                                       ^~~~~~~~~~~~~
test.swift:21:3: note: cannot differentiate through multiple results
  swap(&tuple.0, &tuple.1)
  ^
test.swift:13:40: error: function is not differentiable
print(valueWithGradient(at: 5, 10, in: multiply))
                                       ^~~~~~~~
test.swift:8:11: note: cannot differentiate through multiple results
  let z = tuple(x, y)
          ^

After:

$ swift test.swift
(value: 12.0, gradient: (4.0, 3.0))
(value: 50.0, gradient: (10.0, 5.0))
(value: 12.0, gradient: (4.0, 3.0))
(value: 50.0, gradient: (10.0, 5.0))

[dan-zheng]

Note: the pullback type calculation logic here in VJPEmitter::createEmptyPullback() duplicates SILFunctionType::getAutoDiffDerivativeFunctionType. It should be deduped, I'll look into it soon.

[dan-zheng]

We can add more tests regarding "multiple result indices" (e.g. differentiability witness SILGen tests) when landing TF-1288: relaxing @differentiable and @derivative type-checking to allow original declarations with multiple semantic results.

[dan-zheng]

@swift-ci Please smoke test

[marcrasi]

Exciting!