[AutoDiff] Fix variedness propagation for apply inout arguments.

[dan-zheng]

Propagate variedness from apply argument operands to apply inout arguments
(representing results). apply inout arguments are now correctly marked as
active, triggering non-differentiability errors.

Add ApplyInstBase::getInoutArguments for iterating over @inout and
@inout_aliasable arguments.

Add non-differentiability diagnostics and activity info tests.

Resolves TF-974.

---

Example:

@differentiable
func activeInoutArgNonactiveInitialResult(_ x: Float) -> Float {
  var result: Float = 1
  result += x // Active `inout` argument: `result`.
  return result
}
print(gradient(at: 3, in: activeInoutArgNonactiveInitialResult))

// Before: successful compilation. Prints 0.0.
// After: compile-time error.
//
//     tf-974.swift:1:2: error: function is not differentiable
//     @differentiable
//     ~^~~~~~~~~~~~~~
//     tf-974.swift:2:6: note: when differentiating this function definition
//     func activeInoutArgNonactiveInitialResult(_ x: Float) -> Float {
//          ^
//     tf-974.swift:4:10: note: cannot differentiate through 'inout' arguments
//       result += x // Active `inout` argument: `result`.
//              ^

[dan-zheng]

Note: getInoutArguments is currently defined in the "full application" case of ApplyInstBase. This required duplicating asImpl from the "partial application" ApplyInstBase superclass, which is a private function.

Other options:

• Define getInoutArguments as a top-level function in Differentiation.cpp specialized for just ApplyInst. No problems there.
• Move getInoutArguments to the "partial application" superclass ApplyInstBase. This requires more duplication because getArgumentsWithoutIndirectResults is defined only in the "full application" ApplyInstBase.

[dan-zheng]

Note: @inout-ness is represented in SILParameterInfo. It's quite more complicated to get SILParameterInfo from just an argument Operand (need to use operand index to compute argument index) so I decided to iterate over argument indices.

[dan-zheng]

@swift-ci Please test tensorflow