[MLIR] `scf-parallel-loop-fusion` misses fusion with permuted producer-consumer indices

`scf-parallel-loop-fusion` misses a producer-consumer fusion opportunity between two adjacent `scf.parallel` loops with identical bounds.
The producer writes: `%arg0[i0, i1, i2, i3, i4, i5] = 5.0`. The consumer reads the same buffer with a reversed index permutation: `%v = %arg0[i5, i4, i3, i2, i1, i0]`, and stores it to `%arg1`.

Since both loops iterate over the full same `2x2x2x2x2x2` iteration space, the reversed access is still a one-to-one permutation of the producer iterations. The loops could be fused by mapping each consumer iteration to the corresponding producer iteration. Currently, the loops remain separate, resulting in a missed optimization.

### Input Program

**input.mlir**

```llvm
module {
func.func @permutation_budget(%arg0: memref<2x2x2x2x2x2xf32>, %arg1: memref<2x2x2x2x2x2xf32>) {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%cst = arith.constant 5.000000e+00 : f32
scf.parallel (%arg2, %arg3, %arg4, %arg5, %arg6, %arg7) = (%c0, %c0, %c0, %c0, %c0, %c0) to (%c2, %c2, %c2, %c2, %c2, %c2) step (%c1, %c1, %c1, %c1, %c1, %c1) {
memref.store %cst, %arg0[%arg2, %arg3, %arg4, %arg5, %arg6, %arg7] : memref<2x2x2x2x2x2xf32>
scf.reduce
}
scf.parallel (%arg2, %arg3, %arg4, %arg5, %arg6, %arg7) = (%c0, %c0, %c0, %c0, %c0, %c0) to (%c2, %c2, %c2, %c2, %c2, %c2) step (%c1, %c1, %c1, %c1, %c1, %c1) {
%0 = memref.load %arg0[%arg7, %arg6, %arg5, %arg4, %arg3, %arg2] : memref<2x2x2x2x2x2xf32>
memref.store %0, %arg1[%arg2, %arg3, %arg4, %arg5, %arg6, %arg7] : memref<2x2x2x2x2x2xf32>
scf.reduce
}
return
}
}
```
### Command
```bash
mlir-opt input.mlir --scf-parallel-loop-fusion
```

Version: b1e0aa39763434060ab6e4dc87af3630ac9a04ac

Run the input.mlir example with mlir-opt --scf-parallel-loop-fusion and inspect the scf-parallel-loop-fusion pass to see how producer and consumer indices are matched. Done means the reversed one-to-one permutation is recognized and the two adjacent scf.parallel loops are fused in the output.

Written by the indexing model from the issue text.