Add tanh as a first-class TensorOps activation primitive #630

Problem

tanh is one of the canonical activation functions in ML pedagogy and the default in micrograd, the most-tutorial'd autodiff library on the planet. SKaiNET 0.25.0 ships relu, gelu, elu, leakyRelu, silu, sigmoid, softmax, and logSoftmax on TensorOps — but not tanh.

The gap is visible inside this codebase already: skainet-compile/skainet-compile-dag/.../DefaultExecutionTape.kt:910-914 has the comment:

// Since we don't have tanh yet, we can use: tanh(x) = (exp(2x) - 1) / (exp(2x) + 1)
// or just use a simpler approximation if tanh is missing.
// Actually, let's implement a simple tanh using sigmoid if possible.
// tanh(x) = 2 * sigmoid(2x) - 1

— a local lambda named tanh synthesized inside the GELU backward formula because no real tanh op exists. The downstream skainet-notebook repo has a tanh polyfill doing the same 2*sigmoid(2x) - 1 composition for its micrograd-tribute notebooks. Every consumer re-derives the same expression — that's exactly the kind of duplication a core primitive should eliminate.

Proposal

Add tanh symmetrically with sigmoid. The wiring is the existing six-layer pattern (none of it is novel except the math):

1. TensorOps interface — @Diff @ActivationDsl public fun <T : DType, V> tanh(tensor: Tensor<T, V>): Tensor<T, V>
2. TanhOperation<T, V> : BaseOperation("tanh", "activation", ...) in TensorOperations.kt
3. Tensor<T, V>.tanh() extension in TensorExtensions.kt
4. VoidTensorOps.tanh shape-passthrough
5. DefaultCpuOps.tanh delegating to kotlin.math.tanh (numerically stable across the full float range)
6. RecordingTensorOpsDecorator.tanh records TanhOperation to the tape; stableInputName learns TanhOperation -> "input"
7. Autograd backward: d/dx tanh(x) = 1 - tanh(x)^2, alongside the existing SigmoidOperation backward entry

Full PRD draft (with code shapes per module) is in the downstream repo at docs/upstream/tanh-activation.md.

Release shape

Purely additive — new method on TensorOps, new op class, new extension function. No existing call site changes behaviour. Safe for a 0.25.1 patch release; doesn't require a minor bump.

Downstream consumers waiting on this

• skainet-notebook micrograd demos — the two existing notebooks polyfill tanh and the polyfill deletes entirely once this lands; their rendered graphs collapse from a mulScalar -> sigmoid -> mulScalar -> subScalar chain to a single tanh block.
• skainet-notebook issue for the full micrograd/demo.ipynb port (training + decision boundary) — blocked on the polyfill's perf cost (4× op count per tanh call).
• The GELU backward formula in DefaultExecutionTape.kt:898-919 would collapse from "synthesize a local tanh from sigmoid" to "call ops.tanh(...)" — separate cleanup PR after the primitive lands.

Non-goals

• Don't approximate. kotlin.math.tanh is fast and IEEE-correct; no reason to ship the 2*sigmoid(2x) - 1 composition as the CPU primitive even though it's the polyfill shape.
• Don't add Hardtanh/clamped variants in the same PR (separate op, separate use case).
• Don't change GELU's backward in the same PR — focused cleanup commit after tanh is verified.