fix(qonnx): handle scalar zpt for per-tensor bias quantization

[narutozxp]

When Conv bias quantization is enabled with per-tensor quantization, exported QONNX models may not always use the same shape representation for scale and zero-point. The scale may be stored as (1,), while the zero-point may be stored as ().

The previous logic relied on shape comparison against (1,), which could fail to catch this scalar zero-point representation. If this case is not handled explicitly, the subsequent reshape operation can fail because the zero-point shape is not normalized to the expected per-tensor form.

This PR makes the check semantically correct by treating both forms as valid per-tensor quantization parameters.