feat(nn): Lstm layer with explicit-state step() API

[michalharakal]

Implements the DARC proposal #823 — C (Code) phase.

What

Lstm<T, V> in skainet-lang-core, mirroring Gru's unroll-at-trace-time design, built from
existing primitives only (matmul/narrow/sigmoid/tanh/multiply/reshape/unsqueeze/concat) — no
new TensorOps op, traces to StableHLO with no dedicated converter.

• Gate order i, f, g, o and dual biases match torch.nn.LSTM; weights load after transposing
  to the matmul-ready [in, 4H] / [H, 4H] orientation (same convention as Gru).
• New idiom: LstmState(h, c) + step(xt, state, ctx) — explicit caller-owned recurrent
  state, required by transducer prediction networks (RNN-T/TDT, e.g. Parakeet) and exactly the
  shape that lowers to a fixed-shape single-step StableHLO graph (state as graph I/O).
• initialState(batch, ctx, dtype) helper.

Tests

LstmTest (skainet-backend-cpu, mirrors GruTest):

• scalar-reference parity (all four gates + cell feedback), 1e-5
• step() == unrolled forward() equivalence (the transducer usage pattern)
• gate-order sanity: forget-gate cell persistence
• output shapes

apiDump synced (jvm), apiCheck green.

Motivation / consumers

• Parakeet FastConformer-TDT (SKaiNET-parakeet): 2-layer LSTM prediction network — blocking
• Silero VAD port (SL2610 demo, PLAN 4.1) — the last runtime-Python piece

Closes #823

🤖 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-824 artifact to view the complete documentation locally.

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