[wip] herd runtime parameter support for npu

[fifield]

work in progress to lower air.herd integer operands to mlir-aie runtime parameters. i.e aie.buffer + npu.rtp_write + memref.load

[fifield]

note: comment out last line and uncomment previous to enable, otherwise end up with runtime sequence like,

  aiex.runtime_sequence @mul(%arg0: memref<1024xi32>, %arg1: memref<1024xi32>, %arg2: memref<1024xi32>) {
      aiex.npu.dma_memcpy_nd(0, 0, %arg0[0, 0, 0, 0][1, 1, 2, 512][0, 0, 512, 1]) {id = 0 : i64, metadata = @airMemcpyId4} : memref<1024xi32>
      aiex.npu.dma_memcpy_nd(0, 0, %arg1[0, 0, 0, 0][1, 1, 2, 512][0, 0, 512, 1]) {id = 1 : i64, metadata = @airMemcpyId5} : memref<1024xi32>
      aiex.npu.dma_memcpy_nd(0, 0, %arg2[0, 0, 0, 0][1, 1, 2, 512][0, 0, 512, 1]) {id = 2 : i64, metadata = @airMemcpyId6} : memref<1024xi32>
      aiex.npu.sync {channel = 0 : i32, column = 0 : i32, column_num = 1 : i32, direction = 0 : i32, row = 0 : i32, row_num = 1 : i32}
      aiex.npu.rtp_write(@__air_herd_rtp_0_2, 0, 32)
      aiex.npu.write32 {address = 126976 : ui32, column = 0 : i32, row = 2 : i32, value = 1 : ui32}
    }

where the sync happens before the herd runs (i.e. before the write32 releasing the herd lock)

[fifield]

A better solution is to require the source program to be correct w.r.t. synchronization. Unfortunately mlir-air discards that synchronization information before this point in the lowering, and tries to enforce a synchronization on writes to external memory here instead.