onnx,core: fuse standard ONNX LSTM cell into one LstmEpilogue op

[czoli1976]

Summary

When testing DTLN — an LSTM-heavy real-time speech denoiser — on tract, profiling showed that in streaming mode (batch=1, one audio frame per call) the LSTM body spends most of its non-matmul time on per-gate op dispatch, not arithmetic. tract's ONNX importer decomposes each LSTM cell into ~15 separately-evaluated Sigmoid/Tanh/Mul/Add ops, each materialising an intermediate tensor. The matmuls are already efficient (memory-bandwidth-bound), so this scaffolding is the dominant overhead for streaming LSTMs.

This PR collapses that chain into one fused LstmEpilogue op, emitted from the importer for the standard cell.

What changed

• New LstmEpilogue op (core/src/ops/lstm_cell.rs): given the gate pre-activations [batch, 4*hidden] (ONNX gate order i, o, f, c) and c_prev, computes Ht/Ct in one pass — sigmoid over the i,o,f block + tanh over the c block via tract's vectorised sigmoid_f32/tanh_f32 kernels on contiguous slices, then the cell/hidden update. Replaces the ~15-node decomposed chain.
• Importer emission (onnx/src/ops/rec/lstm.rs): emits LstmEpilogue for the standard cell (no peepholes, concrete hidden size); peephole / symbolic-hidden cells keep the existing decomposed path. tract's ONNX LSTM always uses sigmoid/tanh activations, so the fused op is always valid where it fires. Also computes the gate projections with one matmul per side (Xt·Wᵀ, Ht-1·Rᵀ) + slice rather than four separate matmuls.

Correctness

Same activation kernels as the decomposed path → numerically identical. DTLN end-to-end output stays at 110.47 dB / Pearson 1.00000 vs the native reference (bit-equivalent up to f32 rounding). A unit test checks the cell math against a scalar reference over a multi-row batch.

Perf gain

DTLN, M1 Pro, f32, single-thread, full pipeline (the production regime), restaurant-noise clip, min of 3 runs; lower = faster:

engine	ms / 1 s audio	×realtime
tract main (before)	14.71	68×
tract + this PR	13.11	76×
TFLite — Ruy (TFLite default)	13.80	72×
TFLite — XNNPACK delegate	11.76	85×

• −11% vs tract main (14.71 → 13.11).
• Now ~5% faster than TFLite's default (Ruy) engine (13.80).
• TFLite's XNNPACK delegate (11.76) is still ~12% ahead; that residual is matmul memory-bandwidth plus a couple of streaming-graph follow-ups (below), not this epilogue.

The win concentrates in streaming / small-batch inference, where per-op dispatch dominates the small per-frame matmuls. It is parity-preserving for all shapes; for large offline batches the relative gain shrinks as the matmuls dominate.

Who benefits

Any ONNX model with a standard LSTM run in streaming / small-batch mode — e.g. real-time speech enhancement (DTLN, NSNet), streaming ASR LSTM decoders (RNN-T predictors), TTS (Tacotron-2), OCR (CRNN), time-series forecasting (DeepAR).

Scope / notes

• Measured against vanilla main; orthogonal to in-flight perf PRs (linalg/mmm: cache-adaptive 2D-blocking for the single-thread tile walk #2274 matmul cache-blocking is a different, compounding axis; core/ops/scan: reuse body state across iterations (skip per-timestep plan churn) #2257 Scan iter-reuse is neutral on seq=1 streaming).
• Follow-ups (not in this PR): a companion change inlines the single-iteration Scan for streaming RNNs (a further ~−3% on DTLN, → 12.68 ms/s); and a declutter pass for LSTM cells that were unrolled at export time (some tf2onnx outputs contain no LSTM op, so this importer path never sees them).