Fuse quantizelinear for skip layers using multioutput fusions

Follow up from PR #3782
Ex resnet quantized graph after above PR:

NEW:

q -> conv -> dq -> add -> relu -> q .......... -> q -> conv -> dq -> add -> relu -> q
			   |	                                                    |
			   -> step -> q -----------------------------------------> concat -> conv -> ...

Doing some experimental work, it turns out that we get a slight perf boost from moving the skip-connection quantize op before the step op, and fusing it into the previous conv-pointwise kernel. This should probably be done as 2 steps:

• 1. Write pass for multioutput fusion that can fuse to mlir_quant_convolution_dequantizelinear_dequantizelinear_add_add_relu_quantizelinear_quantizelinear
• 2. Write pass to move q before step when this fusion is possible

These should be done in this order since swapping the order of quantize op and step op is generally not preferable unless doing it for this fusion.