[AutoDiff] Correctly propagate optional adjoint through switch_enum

[asl]

Fixes #74978

[asl]

@swift-ci please test

[asl]

So, the code in question looks like this:

// square2(y:)
sil hidden [ossa] @$s3opt7square21yS2dSg_tF : $@convention(thin) (Optional<Double>) -> Double {
// %0 "y"                                         // users: %14, %3, %1
bb0(%0 : $Optional<Double>):
  debug_value %0 : $Optional<Double>, let, name "y", argno 1 // id: %1
  %2 = metatype $@thin Double.Type                // user: %26
  switch_enum %0 : $Optional<Double>, case #Optional.some!enumelt: bb2, case #Optional.none!enumelt: bb1 // id: %3
...
// %13                                            // user: %26
bb2(%13 : $Double):                               // Preds: bb0
  switch_enum %0 : $Optional<Double>, case #Optional.some!enumelt: bb4, case #Optional.none!enumelt: bb3 // id: %14
...
// %24                                            // user: %26
bb4(%24 : $Double):                               // Preds: bb2
  // function_ref static Double.* infix(_:_:)
  %25 = function_ref @$sSd1moiyS2d_SdtFZ : $@convention(method) (Double, Double, @thin Double.Type) -> Double // user: %26
  %26 = apply %25(%13, %24, %2) : $@convention(method) (Double, Double, @thin Double.Type) -> Double // user: %27
  return %26 : $Double                            // id: %27
} // end sil function '$s3opt7square21yS2dSg_tF'

However, instead of adjoint accumulation (for %0) we just created new adjoint each time. Funny enough, adjoint buffer version for optionals of non-loadable types was correct – adjoint buffers were accumulated together.

Likely the non-optional switch_enum case is broken as well, I just do not have a testcase that directly exposes the problem.

[rxwei]

Nice catch!

[asl]

@swift-ci please test

[asl]

@swift-ci please test