You can do what you want :) in practice it's really hard to bring down the erlang VM in zig. I'm also going to experiment with OTP managed OS threads, which I think should allow you to safely run code that panics (but maybe not segfaults), and will clean up the memory resources after you.
I think in general though you should be using zig not rust, because zig is aggressively unopinionated about memory allocators, so when you allocate memory in your code, you'll use the allocator that zigler provides, which makes the BEAM alloc conformant to zig's allocator interface. This means your code will be more performant, because there are fewer trips to the kernel on allocate and destroy, and your code will respect the BEAM's memory fragmentation, cache optimization, etc.
I love Zig, but it doesn't have the security guarantees (yet) of Rust, so I would recommend either running it out of proc or within an in-proc Wasm environment, which actually has other interesting properties.
edit
https://www.reddit.com/r/elixir/comments/e955jw/zigler_zig_n...
https://hexdocs.pm/zigler/Zigler.html
So this is an Erlang Nif [3] which runs inside the VM, akin to JNI in Java. This also means that errant code can bring down the entire VM.
Would you be open to a fork of Zigler that targeted a Wasm environment running inside the Erlang env?
[3] http://erlang.org/doc/man/erl_nif.html