Do not serialize-deserialize module before calling aie2xclbin

[newling]

This PR does a few things, happy to split it up if requested.

1. Before pulling mlir-aie into iree-amd-aie, it we required to serialize-deserialize to call the aie2xclbin program. But we now just use aie2xclbin as a function, no shell out anymore. So no serialization-deserialization needed.

2. This PR moves dma-to-npu closer to lower-to-aie pass. I think we'd eventually like to change the lowering of npu instructions from

amdaie dialect -> aie dialect -> npu

to

amdaie dialect  -> npu

, because the amdaie and aie dialects are very similar and this indirection doesn't provide us with anything afaict. Making that change is currently not possible (dma-to-npu must currently run after stateful transform pass), this change is a step in that direction though

Test changes: I removed some CHECKs for aiex.runtime_sequence in tests/samples, because that's now sucked into the LX instructions (sequence of integers). In my mind the tests in tests/samples are only useful to check that compilation doesn't error/crash, so IMO removing CHECK lines there is fine.

3. General clean-up, for example we don't need aiex-to-standard pass anymore.

[makslevental]

Currently dma-to-npu runs in aie2xclbin, we shouldn't be running mlir passes there because it's on the other side of the serialization wall.

100% agreed. I recommend you xfail (for a moment) those failing tests to see if this will work e2e.

[makslevental]

Overall looks good but a couple of small questions/nits.

[makslevental]

Cool thanks for this change (it's a small but important one). Sorry for the hassle :(