Once you get to know Lua tables and then metatables, you will question the existence of different data structures :)
One reason using global variables is bad practice is it makes testing harder. An unfortunately high percentage of embedded software doesn't have any sort of harness-based testing because it's written with globals spammed everywhere, which prevents you from using any kind of principled testing strategy where you mock out all the hardware dependencies. It's especially bad if there's globally defined MMIO stuff like "#DEFINE CCR1A (uint64_t)0x74FEA10". Good luck testing that!
In domain of C, passing by reference means passing pointer. If you chuck everything into single struct and pass by pointer, it has same problems as global scope.
Not that I'm advocating for global variables. Even tiny projects tend to grow with time, and localizing scope across the code base is not fun at all. In context of Lua, I've just trained myself to prefix variables with 'local' and I don't give it much thought.
Not true. In fact, this refactor is one of the best things you can do to improve an old shitty embedded C codebase. Among other benefits, it allows you to have multiple instances (an arbitrary and easily-adjusted number, in fact) of a system sharing the same memory, reduces the complexity of linker-related BS, and simplifies testing. It's vastly better than relying on horrendous C cross-module scoping rules for sharing.
I agree that passing in structs is vastly better than communicating over globals. On the other hand, taking existing code base and implementing this state-passing is quite large undertaking that affects all function declarations/implementations. It might be beneficial, but there are often better investments of your time.