Symmetric Transfer in Coroutines

[ashvardanian]

I believe the performance of the coroutine implementation in the benchmark can be significantly improved, to at least match the throughput of experimental coroutines in Rust. The current implementation is relatively minimal, uses std::suspend_always, and implements no Symmetric Transfer.

Changing those can help accelerate the benchmarked pipeline. To learn more about coroutine implementations, the Deciphering C++ Coroutines talk by Andreas Weis from CppCon 2022 is a good start.