onnx: auto-enable external_state to inline single-iteration Scan for streaming RNNs

[czoli1976]

Summary

Companion to #2294, from the same DTLN profiling. tract already has a declutter_single_loop pass that inlines a single-iteration Scan, but it only fires when Scan::external_state is set — which until now was reachable only via the manual force_scan_external_state transform. So streaming RNNs (seq_len == 1, one frame/token per call) carried a dead one-iteration Scan on every call: pure orchestration overhead (state spawn, per-iter setup, output assembly).

This PR has the ONNX recurrent importer set external_state automatically when the caller manages state.

What changed

In onnx/src/ops/rec/common.rs, when an LSTM/GRU/RNN exposes its full recurrent state both as input and output — initial_h + Y_h (and initial_c + Y_c for LSTM) — set Scan::external_state = true. That is the caller-managed-state contract (streaming denoisers, autoregressive / RNN-T decoders): the caller supplies state in and reads it out every call. The existing declutter_single_loop then inlines the seq_len == 1 Scan, feeding the body's State input directly from the outer input each call.

13 lines; no new op or pass — it just enables the existing inliner for the case it was designed for.

Correctness / soundness

external_state only affects declutter_single_loop (the inline guard); it is inert for multi-iteration Scans, so full-sequence (seq_len > 1) runs are unchanged. Inlining a single-iteration Scan with externally-supplied state is sound — the body runs once and state flows through the model I/O (this satisfies the guard from issue #2157).

Measured on DTLN end-to-end: output stays 110.47 dB / Pearson 1.00000 vs the native reference. Existing tract-onnx tests pass.

Perf gain

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

config	ms / 1 s audio	×realtime
tract main	14.71	68×
tract + this PR	13.54	74×
tract + #2294 + this PR	12.68	79×
TFLite — Ruy (TFLite default)	13.80	72×
TFLite — XNNPACK delegate	11.76	85×

• −8% vs main (14.71 → 13.54), edging past TFLite's default (Ruy) engine.
• Stacks with onnx,core: fuse standard ONNX LSTM cell into one LstmEpilogue op #2294 (the LstmEpilogue gate-epilogue fusion): together −14% (→ 12.68), comfortably past Ruy. The two are orthogonal — this PR removes Scan orchestration, onnx,core: fuse standard ONNX LSTM cell into one LstmEpilogue op #2294 fuses the gate epilogue.
• TFLite's XNNPACK delegate (11.76) is still ahead; that residual is matmul memory-bandwidth, not Scan/epilogue overhead.

Who benefits

Every seq_len == 1 streaming RNN exported with exposed state — real-time speech enhancement (DTLN, NSNet), streaming ASR LSTM/GRU decoders (RNN-T predictors), autoregressive TTS, and similar on-device / real-time workloads.

[kali]

So i have tried this before, but ran into a problem. It don't think it is possible to distinguish reliably a stateful model meant to be called with pulses of length 1 from a "stateless" (from tract POV) model with state managed on the outside. The difference is only manifest once the user call it with Runnable::run or ::spawn then State::run, but we need to apply the Transform before that. (Now that I think of it, this is how 0.23 where breaking DFN3, right ?)

WDYT ?

[kali]

Ha, your heuristics is, C and H are outputted so the state is managed externally ?

[czoli1976]

You are right, it has an hole now, as it is

Trying to refine to see if can make it more robust, let's see