Cache callbacks for long running processes

[essen]

See erlang/otp#5831 for details as to why it's faster to do that. Might not be worth doing in many parts of Cowboy, but some like Websocket or the loop handlers may benefit from this.

[geeksilva97]

I'd like to give it a try. Sounds like a nice challenge

[essen]

The hard part is figuring out how to measure the improvements (or lack thereof). It might not be worth outside of receiving Erlang messages (the network is likely slower).

[geeksilva97]

Maybe we could measure the execution time with those Erlang profiling tools?

[essen]

Profiling tools are useful but the deciding factors for performance improvements in Cowboy are latency and throughput so for Websocket for example this means being able to make the clients either receive data faster than before, or receive more data than before, for specific workloads. Currently the only such measurements we have are the run times of the Autobahn test suite.