Implement synchronous SDK client

[spetz]

There might be specific use cases, when making use of the synchronous client yields a better performance than with the async one.
For example, when building the ultra-low latency system, with threads pinned to the cores - the async client is pretty much useless in such a design and will result in worse performance than the sync one.

The sync client could be kept behind a specific feature flag, which would drop the existing Tokio reference, and use the standard TCP network library (along with the blocking client for reqwest crate). The existing Client trait could have all the methods defined as the synchronous ones. The other solution, would be to introduce some sort of SyncClient trait, as an addition to the already existing asynchronous version.

It would also be beneficial, to provide the similar high-level builders for IggyProducer, IggyConsumer` etc. to provide a great developer experience.

This is certainly a big feature, and it's worth discussing how to best tackle this.

[numinnex]

During discussion on discord together with @marvin-hansen we came to conclusion, that the sans-io approach is probably best fit solution for this problem. It's non trivial, but the payoff should be worth it.

This approach would make our client not only async runtime agnostic, but also network transport protocol agnostic aswell.

I link some scraps of knowledge about this approach that I've managed to find during my initial research:

• Cloudflare quiche crate (QUIC implementation using sans-io)
• fasterthanlime video "case for sans-io"

[marvin-hansen]

Great Write up, thank you @numinnex

Just a few links I want to add.

The sans Io blog post

https://www.firezone.dev/blog/sans-io

Let's add the socket factory to the idea list because it's a terrific solution to the problem of setting up and configuring connections for various protocols. That simplifies the SDK further and makes each protocol implementation a bit more uniform which is usually easier to maintain.

https://github.com/firezone/firezone/blob/main/rust/connlib/socket-factory/src/lib.rs

Because the sans-io mechnism if fundamentally effect-free, this is akin to the free-monad so a bit of background for those who missed category theory.

https://okmij.org/ftp/Computation/free-monad.html

There is a pretty bonkers SANS-I/O runtime runtime that is work in progress. I say its bonkers because by their own benchmark, they maxed out at 18GB/sec before proper optimizations..

https://github.com/8xFF/sans-io-runtime