Ground truth: decide on gradient validation (currently fully unimplemented) #986

Sub-issue of #984.

Despite the "gradienttracer" name, no gradient is ever actually checked today.
ToleranceConfig.GRADIENT, GroundTruthTestCase.expectedGradients, and grad_*-tensor
parsing all exist on the Kotlin side (GroundTruthLoader.kt,
GroundTruthValidator.validateGradientsInternal) — but store_experiment_as_gguf
(gradienttracer's gt/pytorch/io/writer.py) never calls .backward() or writes any
grad_* tensor. validateGradientsInternal is explicitly a placeholder (// TODO: Implement actual gradient computation via tape) that always reports failure if ever
invoked with real data.

Decision needed

• Implement it: add .backward() + grad_* tensor serialization to
  store_experiment_as_gguf (every @Executable input already has requires_grad=True
  set, so the gradient is available, just not captured), and implement SKaiNET-side
  backward-pass execution in GroundTruthValidator.validateGradientsInternal to actually
  compare against it. This is real, nontrivial work on both sides — SKaiNET's own
  autodiff/backward execution path needs to be driveable from
  skainet-test-groundtruth, which it currently isn't.
• Explicitly de-scope it: if gradient validation isn't a near-term priority, say so
  in CONTRACT.md and TESTING.md (which currently reads as if this already works),
  and consider whether the dead expectedGradients/ToleranceConfig.GRADIENT/
  validateGradients parameter surface should be removed rather than left as
  looks-implemented-but-isn't.

Either answer is fine — the point of this issue is picking one, since right now the code
reads as "supported" when it silently isn't.