[Pallas] Fix fori_loop multi-dim inner loop index unflattening

[thcmbs]

Fix out-of-bounds DMA access when fori_loop iterates over multiple inner block dimensions

The flat _j loop variable was used directly for all dimensions' offsets; now we unflatten it into per-dimension indices (_j_0, _j_1, ...) via divmod decomposition

[norx1991]

FYI, there are some comments under #1917. We see slightly better performance with double fori_loop, and it also seems a more natural way to implement this.

[thcmbs]

Thanks for sharing! Given the perf diff, there is no debate :)

[thcmbs]

Closing given that #1917 was merged.