ONNX importer: LSTM, Gather, Unsqueeze/Squeeze, Cast, Where, If (Silero VAD import as the gate)

[michalharakal]

Context

skainet-io-onnx's graph importer (OnnxImport.kt) handles Add, BatchNormalization, Concat, Conv, Div, Gemm, LeakyRelu, MatMul, MaxPool, Mul, Relu, Reshape, Resize, Sigmoid, Silu, Slice, Softmax, Split, Sub, Transpose, Upsample — the YOLO-shaped set from #230. Recurrent and control-flow nodes are absent, so common small audio/speech graphs (Silero VAD, RNN-T prediction nets, NeMo exports) cannot be imported even though the DSL has Lstm (#823) and gather.

This is the verification path for the native Silero VAD issue: importing the upstream graph and comparing against the hand-written module catches drift in either.

Scope

• LSTM (ONNX opset ≥ 7 semantics: W, R, B layouts, initial_h/initial_c, direction=forward first; bidirectional as a follow-up) lowered onto Lstm.step unrolled over the sequence axis.
• Gather, Unsqueeze, Squeeze, Cast, Where, Shape/ConstantOfShape as needed by the Silero graph.
• If with constant-foldable conditions (Silero's sample-rate branch) — fold at import when the condition is an initializer, reject otherwise with a clear message rather than silently importing one branch.
• Import test on the pinned silero_vad.onnx: graph builds, forward matches ORT golden within 1e-4 (same golden as the native Silero issue).
• Ground-truth coverage for the new ops (Ground truth: expand Python-side op coverage to match OperationExecutor #985).

Acceptance

• OnnxLoader imports Silero VAD end to end without hand edits to the file.
• Each new converter has a unit test with a minimal one-node graph.

Related

• ONNX I/O module MVP: #230 — ONNX I/O module MVP
• [Feature]: LSTM layer (Lstm + explicit-state step API) in skainet-lang-core #823 — Lstm
• Ground truth: expand Python-side op coverage to match OperationExecutor #985 — ground-truth op coverage

[michalharakal]

Tracked in #1282.