The air-dma-to-channel pass is now more correctly preserving the input dependency, leading to more correct dependency graph but at the cost of mroe bds. Test modified to limit the number of bds per queue

Author: erwei-xilinx

test/xrt/11_gemm_bias_fusion/gen.py

@@ -190,5 +190,6 @@
         air_loop_fusion=True,
         omit_auto_broadcast=True,
         channel_multiplexing=["L1",],
+        runtime_loop_tiling_sizes=[2, 2],
     )
     backend.compile(air_module)