[Lane 2 · kotlin-core] Precision/Recall/F1 — extract Accuracy's class-matching helpers into a shared internal file #1225

Sub-issue of #1222 (Precision, Recall and F1Score metrics).

Lane: 2 · Kotlin core (preparatory refactor)
Skill needed: Kotlin. No tensor-internals knowledge needed — everything you move already exists and already has tests.
Size: s (~2–3 h)
Blocked by: nothing. Unblocks #1226.

What to do

Accuracy.kt (skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/nn/metrics/Accuracy.kt)
carries, as private members, all the class-matching machinery the three new metrics also
need: argmax, argmaxTarget, insertClassIndex, validateIndexTargetShapes,
validatePredictions, normalizeDim, iterateOverSamples, plus the "hard Int32 targets
vs. soft FP32/FP16 targets" dispatch and the binary-threshold branch in
computeWithIndexTargets / computeWithSoftTargets. Copying that three more times is the
duplication git blame will ask about in six months.

1. Create ClassificationSupport.kt in the same package. Move the helpers there as
   internal top-level functions (or one internal object ClassificationSupport). Keep
   names and signatures; keep error messages, but make the "Accuracy …" prefixes
   metric-agnostic (pass the metric name in, or drop it).
2. Add one internal entry point that does the per-sample walk once for both target
   kinds and hands (predictedClass, targetClass) pairs to a callback, e.g.
   internal inline fun forEachPrediction(predictions, targets, dim, threshold, action: (predicted: Int, target: Int) -> Unit).
   Accuracy.update becomes a ~5-line call to it that counts matches. This is what
   ConfusionMatrixAccumulator ([Lane 2 · kotlin-core] Precision/Recall/F1 — Averaging enum + internal ConfusionMatrixAccumulator #1226) will call too.
3. Accuracy's public surface must not change: same constructor, same accuracy() /
   binaryAccuracy() factories, same behaviour. Run
   ./gradlew :skainet-lang:skainet-lang-core:apiCheck — the API dump must be unchanged
   (everything you add is internal).
4. Run the existing tests unchanged:
   ./gradlew :skainet-backends:skainet-backend-cpu:jvmTest --tests "sk.ainet.exec.metrics.*".

Acceptance

• Accuracy.kt contains no argmax / index-walking code of its own
• apiCheck passes with no dump change
• AccuracyTest passes unmodified
• PR title refactor(metrics): extract classification helpers from Accuracy (#1222)

Notes

Pure refactor — no new behaviour, so no new tests needed beyond the existing ones passing.
If you find Accuracy has a latent bug while you're in there (for example the
threshold branch reads logit index 1 when classCount == 2 but index 0 when
classCount == 1 — check that this is intended), don't fix it here; open a separate
bug and link it on the parent.