Both ways seem possible.
What is the better approach within Linux? Better means more efficient assuming my switch doesn't know about LACP+Tagging.
You can do either with recent kernels, there is no appreciable difference in efficiency.
Personally I find the idea of putting a VLAN interface into a bond to be weird, if you want high availability then surely you want it for the whole physical interface then sort out VLAN tagging later, but enough people asked for it that it's a thing now.