Add correct int type to the quantized weights matrix in MatMulInteger

[peterbohm]

Assignment of the quantized weights to B matrix in MatMulInteger, was casting them to the same int type as the quantized input (uint8).
This created an issue for models with weights quantized to int8 (e.g. using onnxruntime.quantization.quantize_dynamic()).
This commit checks for the actual data type of the weights and then uses that during assignment to B.

[kraiskil]

Thanks for the PR.

ONNX specifies the two inputs to be of different types, so this is a clear bug in onnx2c. Good to merge.

[kraiskil]

Thanks :)