[mlir][linalg] Fix interchange initialization in fusion on tensors.

[gysit]

Summary

If no interchange vector is given initialize it with the identity permutation from 0 to number of loops.

mlir/lib/Dialect/Linalg/Transforms/FusionOnTensors.cpp

void runOnFunction() override {
     }
 
     // Copy the `tileSizes` and `tileInterchange` prefixes needed to tile
     // `rootOp` or use the identity interchange if `tileInterchange` is empty.
     SmallVector<int64_t> rootTileSizes(
         tileSizes.begin(), tileSizes.begin() + rootOp.getNumLoops());
     SmallVector<int64_t> rootInterchange =
         tileInterchange.empty()
-            ? llvm::to_vector<6>(llvm::seq<int64_t>(0, tileSizes.size()))
+            ? llvm::to_vector<6>(llvm::seq<int64_t>(0, rootOp.getNumLoops()))
             : SmallVector<int64_t>(tileInterchange.begin(),
                                    tileInterchange.begin() +
                                        rootOp.getNumLoops());
 
     // Check `rootInterchange` is a permutation of the `rootOp` loop dimensions.
     // It has to be a permutation since the tiling cannot tile the same loop
     // dimension multiple times.
     if (!isPermutation(rootInterchange))