[OptRed] Extend -tritonintelgpu-optimize-reduction-locality to support repCluster[0] > 2 (#2533)

Support `repCluster[0] > 2` by using 7-D tensors and adding a
`convert_layout` operation before the final `reshape`.

See code for implementation details.

---------

Signed-off-by: victor-eds <[email protected]>

Author: victor-eds