Refactor AIRLowering pass with mlir typeConverter

[erwei-xilinx]

No description provided.

[fifield]

doesn't it already contain air-dma-to-channel? Do you just need a different order?
Starting about line 375:

        pass_pipeline = ",".join(
            [
                "air-insert-launch-and-segment-around-herd",
                "func.func(air-lower-herd-parallel)",
            ]
            + (
                EXPERIMENTAL_PASSES
                if "npu" in opts.device and opts.experimental_passes
                else []
            )
            + (["air-dma-to-channel"] if "npu" in opts.device else [])
            + [
                "canonicalize",
                "cse",
                "air-specialize-channel-wrap-and-stride",
                "func.func(air-renumber-dma)",
                "func.func(convert-linalg-to-loops)",
                air_place_pass,
            ]

[erwei-xilinx]

Yes we need air-dma-to-channel to produce us with air.channel ops before the following passes:

    "func.func(air-loop-fusion)",
    "air-label-scf-for-to-ping-pong",
    "air-ping-pong-transform{keep-memref-dealloc=true}",
    "air-isolate-async-dma-loop-nests",

Those passes are for BD optimizations, and they only work on air.channels, because air.dma_memcpy_nd doesn't have explicit handles for BDs.