[dequantizeLinear] ONNX only supports one axis dequantizeLinear

[lisa0314]

Consider to add emulation to support multiple axes case, e.g. input shape is [2, 3, 4, 5] and scale shape is [1, 3, 4, 1].

[fdwr]

Note blockwise broadcasting via emulation is just a ONNX resampling call (https://onnx.ai/onnx/operators/onnx__Resize.html) with mode "nearest" (treat as nop if there isn't any scaling).

function dequantizeLinear(input:uint8, scale:float32, zeroPoint:uint8)
    upsampledScale = resample(scale, sizes=input.shape)
    upsampledZeroPoint = resample(zeroPoint, sizes=input.shape)
    return mul(sub(input, upsampledZeroPoint).cast(float32), upsampledScale)
endfunction