Is the shared memory implementation worth unsafe?

[AaronFriel]

A significant amount of unsafe code in the repository is providing an implementation for shared memory between the AICI runtime and the LLM engine.

It would be good to:

1. Benchmark the current approach, understand latency
2. Compare to alternatives that are more ergonomic (zeromq IPC, gRPC or Apache Arrow Flight¹)
3. If syscalls are remain too high overhead, evaluate moving the shared memory protocol to a crate that provides only a safe interface and use miri, fuzzing.

Footnotes

1. Which may eventually support shared memory transports for Rust and Python clients. ↩