Missing fusion for winograd transform ops with their consumers

[Han-Chung Wang]

This is very helpful in the winograd + demotion path. In the winograd transformation, a convolution becomes (winograd transform ops -> batch_matmul -> winograd_output_transform). After demotion, there would be a turncf op right after winograd input/filter transform ops. We want to fuse them, as it is very straight-forward.

I think we can use this example to start more LinalgExt fusion work.

(cc @IanWood1 @MaheshRavishankar who recently works on LinalgExt fusion)