[DmaLoopSubsumption] Fix for strided op in loop without induction var dependency

[jtuyls]

Fixes an issue exposed by a 128x32x64 matmul: #556.

This issue would come up in partial loop dependencies in case of scf.forall:

scf.forall (%arg2, %arg3) in (2, 6) {
  %1 = amdaie.npu.dma_cpy_nd %0([0, %arg3] [8, 16] [16, 1], [] [] [])
  amdaie.npu.dma_wait(%2, S2MM)
}

In this case, the outer iteration upon which the NPU DMA operation didn't depend ( %arg2) would not be considered, resulting in incorrect output IR:

%1 = amdaie.npu.dma_cpy_nd %0([0, 0, 0] [6, 8, 16] [1, 16, 1], [] [] [])
amdaie.npu.dma_wait(%2, S2MM)

However, the outer iteration should be considered as well, making sure the above DMA operation is executed twice, resulting in the below output IR:

%1 = amdaie.npu.dma_cpy_nd %0([0, 0, 0, 0] [2, 6, 8, 16] [0, 1, 16, 1], [] [] [])
amdaie.npu.dma_wait(%2, S2MM)

This PR fixes the issue by adding general support for subsuming loop iterations into strided operations without any loop induction variable dependency.

[nirvedhmeshram]

Nice fixes, LGTM!

[Abhishek-Varma]

LGTM!

[jtuyls]

I noticed only the outermost dimension can be zero in HW, so I will create a follow-up to not perform the loop subsumption in that case. As that is a specific HW limitation, I think it's not needed to alter the general behaviour of this transformation for that, so will use a flag to enable/disable that limitation.