Add GRU layer (composed, trainable, StableHLO-exportable)

[michalharakal]

Closes #217.

What

The first recurrent layer in SKaiNET. GRU is implemented as a Module composing existing primitive ops, unrolled over the static sequence length at trace time.

Why composed rather than a fused op + dedicated converter: StableHLO export has no control flow (no while/scan), so a recurrence must unroll regardless of approach. A composed layer therefore emits the same unrolled primitive graph a fused convertGru would — but reuses every existing, already-validated converter (matmul/add/sigmoid/tanh/multiply/narrow/concat) with zero converter changes, and is trainable for free because every one of those primitives already has a correct autodiff backward.

Scope (first cut)

Single-layer, unidirectional, batch-first: [B, S, D] → [B, S, H]. PyTorch-compatible gate math and order (reset, update, new); weights stored matmul-ready so torch.nn.GRU weights load after a transpose:

r = sigmoid(x·Wir + b_ir + h·Whr + b_hr)
z = sigmoid(x·Wiz + b_iz + h·Whz + b_hz)
n = tanh   (x·Win + b_in + r ⊙ (h·Whn + b_hn))
h' = (1 - z) ⊙ n + z ⊙ h

Bidirectional / num_layers / packed sequences are explicit future work.

Changes

• nn/Gru.kt — the layer (Module + ModuleParameters; params weight_ih/weight_hh [*,3H], bias_ih/bias_hh [3H]), unrolled forward.
• nn/dsl/NetworkBuilder.kt — GRU dsl item + GruImpl + gru(hiddenSize){} builder (input size inferred from the preceding layer).

Tests

• GruTest — eager forward vs an independent scalar reference (all gates + recurrence + hidden feedback).
• ConvPoolBackwardTest.gru_backward_input_matches_finite_diff — input gradient vs central finite difference (proves grads flow through the unrolled cell).
• GruDslTest — network { gru(8) } wires a Gru module.

All three modules' suites pass locally (skainet-lang-core, skainet-backend-cpu, skainet-compile-dag).

Follow-up

End-to-end StableHLO → IREE export validation is added in the conformance suite (gated on a release carrying this layer), same flow as the norm/upsample ops.

🤖 Generated with Claude Code

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-772 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.