feat: cross-runtime quantized comparison (pytorch-quantized vs onnx-quantized)

[SAY-5]

v4: cross-runtime ONNX comparison

Exports each PTQ config to ONNX (FP32 via torch.onnx.export, INT8 via onnxruntime.quantization) and benches under ONNX Runtime CPU EP. Compares top-1 / latency / on-disk size against the PyTorch quantized runtime. Asserts top-1 parity within +/-1pp.

Numbers (full 10k CIFAR-10 test split, M-series CPU):

config	pt_top1	onnx_top1	top1_pp	pt_p50_ms	onnx_p50_ms
fp32_baseline	82.3%	82.3%	0.00	1.83	0.83
dynamic_int8	82.3%	82.3%	0.00	1.14	0.38
static_int8_per_tensor	82.1%	82.1%	-0.05	1.77	0.18
static_int8_per_channel	82.0%	82.3%	+0.27	1.27	0.18

All four configs pass the +/-1pp structural-parity gate. See artifacts/results/cross_runtime.{json,md} and docs/cross_runtime.md.