Fuse quantized matmul subgraphs for dynamic quantization

[robertknight]

Quantizing a model using dynamic quantization will produce subgraphs like this for quantized matmuls:

In text:

Y, Y_zero, Y_scale = DynamicQuantizeLinear(X)
Y = MatMulInteger(Y, W_quant, Y_zero, W_zero)
Y = Cast(Y, to: f32)
Z = Mul(Y_scale, W_scale)
Y = Mul(Y, Z)

Where the W_* values are constants.

The Cast here can obviously be fused into the MatMulInteger op, as could the final Mul op of the float matrix by Z scale vector.