Offset and validation fixes for npu_memcpy_nd

[andrej]

1538 introduced a bug (#1578) to the offset calculation in npu_dma_memcpy. This was not caught because there were not sufficient tests for this feature. This PR

• fixes the erroneous offset bug
• refactors the stride/wrap/offset calculations into their own functions; this way they can be reused for validation (we want to validate the same values that we will later lower)
• adds more validation code for npu_dma_memcpy, to make sure input strides/wraps/offsets are valid (instead of just rounding to the nearest valid value, which is probably almost never what the user expects)
• adds more test cases for npu_dma_memcpy

[andrej]

Thanks for having a look @fifield and @AndraBisca. I think I've addressed all feedback