[mlir] Fix canonicalization of tiled_loop if not all opresults fold.

The current canonicalization did not remap operation results correctly
and attempted to erase tiledLoop, which is incorrect if not all tensor
results are folded.