AIRSegmentLoopFusion: A number of fixups and improvements around async dependency

[erwei-xilinx]

• Trace async token users through air.wait_all;
• Move fused loop to just before the last loop being fused, to avoid any ssa dominance issue;
• More informative failure message when broken dependence between air.channel.get---as data producer---and air.channel.put---as data consumer---is detected in the generated fused scf.for loop.