Miopen dialect opt step18 : add source_bounds and output_bounds attributes in miopen.transform

[whchung]

Added a custom builder API for miopen.transform to populate source_bounds and output_bounds attributes in miopen.transform.

when we transform A to B in miopen.transform, the dimension of B becomes source_bounds, the dimension of A becomes output_bounds, because A would be lower level tensor and B would be the upper level tensor. when we apply coordinate transformations we begin from upper level coordinates.

this is also an infrastructure need for F_infinite algorithm to properly compute index diff maps for semi-affine transformations such as Merge.

[asroy]

when we transform A to B in miopen.transform, the dimension of B becomes source_bounds, the dimension of A becomes output_bounds, because A would be lower level tensor and B would be the upper level tensor. when we apply coordinate transformations we begin from upper level coordinates.

Maybe we should call A/B lower/upper instead of source/output, so we don't need to switch between two opposite naming scheme when we switch between transforming descriptor and transforming coordinate

[whchung]

when we transform A to B in miopen.transform, the dimension of B becomes source_bounds, the dimension of A becomes output_bounds, because A would be lower level tensor and B would be the upper level tensor. when we apply coordinate transformations we begin from upper level coordinates.

Maybe we should call A/B lower/upper instead of source/output, so we don't need to switch between two opposite naming scheme when we switch between transforming descriptor and transforming coordinate

Will file it as a separate PR.

[whchung]

when we transform A to B in miopen.transform, the dimension of B becomes source_bounds, the dimension of A becomes output_bounds, because A would be lower level tensor and B would be the upper level tensor. when we apply coordinate transformations we begin from upper level coordinates.

Maybe we should call A/B lower/upper instead of source/output, so we don't need to switch between two opposite naming scheme when we switch between transforming descriptor and transforming coordinate

Please refer to #243 for such an implementation.