[Dma] Don't check Dma unit dims during subsumption and fold them during controlCodeLowering

[yzhang93]

Before this PR, we need to check the number of sizes/strides before dma loop subsumption to make sure the number of dims are not exceed the maximum after subsumption. However, this blocks some opportunities for loop subsumption when there are unit dimensions which are not canonicalized because the offsets of these dimensions are none 0. For example the following loop cannot be subsumed because there are already 4 dimensions on L3 source side.

scf.for %arg2 = %c0 to %c6 step %c1 {
  %1 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
  amdaie.npu.dma_cpy_nd %0([] [] [], [1, %1, 0, 0] [1, 1, 32, 32] [8192, 1024, 32, 1])
}

This PR relaxes the constraint by only checking the non-unit dimensions, so the above loop can be subsumed into dma as

amdaie.npu.dma_cpy_nd %0([] [] [], [0, 1, 1, 0, 0] [6, 1, 1, 32, 32] [1024, 8192, 1024, 32, 1])

And this dma can be further canonicalized.

[jtuyls]

LGTM