fix: prevent panics in shape inference when operands missing

[jaweed3]

Problem

Five ops (Concat, Squeeze, Gemm, Reshape, and the RNN family) index inputs[k] / outputs[k] in their rules() without arity validation. A malformed ONNX with too few operands triggers an index-out-of-bounds panic instead of a graceful error.

Fix

Add check_input_arity / check_output_arity guard at the top of each rules() before any indexing.

Op	File	Guard
Concat	hir/src/ops/array/concat.rs	check_input_arity(inputs, 1)?;
Squeeze	onnx/src/ops/array/squeeze.rs	check_input_arity(inputs, 1)?;
Gemm	onnx/src/ops/math/gemm.rs	check_input_arity(inputs, 2)?;
Reshape	hir/src/ops/array/reshape.rs	check_input_arity(inputs, 2)?;
RNN	onnx/src/ops/rec/common.rs	check_output_arity(outputs, output_count.max(1))?;

Verification

Before: tract concat_poc.onnx dump → panic at concat.rs:37
After: tract concat_poc.onnx dump → Error: Wrong input number. Rules expect 1, node has 0.

Closes #2391.

[github-actions]

🟡 Bench vs main — no speed regressions · 1 load/memory mover(s)

Reference: main nightly, latest 2026-06-25 (0d old) · PR ddadf4c0d · ran on apple-m1-max, i9-11900kb_rtx-4060, jetson-orin-nx · 651 metrics compared

Speed — evaltime · prefill · decode

no inference-speed regressions

Load & memory (worst first)

Δ	metric	device	main → PR
🔴 +6.1%	openelm_270M_q40ef16_541 RSS @ ready · metal	apple-m1-max	299 MB → 318 MB

🟢 3 improvement(s)

Δ	metric	device	main → PR
🟢 -26.7%	parakeet_tdt_600m_v3_f32f32_decoder_pass load+optimize · metal	apple-m1-max	146 ms → 107 ms
🟢 -6.2%	llama_3_2_3B_q40ef32_516 load+optimize · metal	apple-m1-max	1.56 s → 1.46 s
🟢 -5.1%	llama_3_2_3B_q40ef32_516 load · metal	apple-m1-max	1.36 s → 1.29 s

lower is better except prefill/decode (tok/s) · adaptive thresholds (max(floor, k×noise) vs the series' own history) · single-shot vs nightly reference · full report → run