Fuse "contiguous + concat"

@212 = reshape_lazy[dims={1, 256, 20, 1, 20, 1}](@210) -> half_type, {1, 256, 20, 1, 20, 1}, {102400, 400, 20, 20, 1, 1}, target_id=0: 0.00093558ms, 1%
@213 = load[offset=4505600,end=5324800](@1) -> half_type, {1, 256, 20, 2, 20, 2}, {409600, 1600, 80, 40, 2, 1}, target_id=0: 0.00062104ms, 1%
@214 = multibroadcast[out_lens={1, 256, 20, 2, 20, 2},out_dyn_dims={}](@212) -> half_type, {1, 256, 20, 2, 20, 2}, {102400, 400, 20, 0, 1, 0}, target_id=0: 0.00112586ms, 1%
@215 = gpu::code_object[code_object=9168,symbol_name=contiguous_kernel,global=204800,local=1024,](@214,@213) -> half_type, {1, 256, 20, 2, 20, 2}, {409600, 1600, 80, 40, 2, 1}, target_id=0: 0.0184781ms, 1%
@216 = load[offset=2867200,end=4505600](@1) -> half_type, {1, 512, 40, 40}, {819200, 1600, 40, 1}, target_id=0: 0.00066402ms, 1%
@217 = reshape_lazy[dims={1, 256, 40, 40}](@215) -> half_type, {1, 256, 40, 40}, {409600, 1600, 40, 1}, target_id=0: 0.00062264ms, 1%
@218 = gpu::code_object[code_object=9352,symbol_name=concat_kernel,global=204800,local=1024,](@217,@150,@216) -> half_type, {1, 512, 40, 40}, {819200, 1600, 40, 1}, target_id=0: 0.0205865ms, 1%

This pattern is from YOLOv5s model.

For this case :
@150 which is the second input to the concat can be reshaped to {1, 256, 20, 2, 20, 2} and then can be concatenated with @215 which is the ouptut of the contiguous.

and then concat output can be reshaped to {1, 512, 40, 40}.

This way contiguous + concat can be fused togther.

Start by locating the graph optimization entry points for contiguous and concat, then reproduce the YOLOv5s pattern shown by nodes @150, @215, and @218. Confirm that the contiguous operation is fused into concat and that the final output has shape {1, 512, 40, 40}; the issue names no files or tests, so those must be found in the repository.