$ref keyword caches built validators

[macisamuele]

The goal of this PR is to reduce validation penalty performance while interacting with $refs.

At the current state every time we're trying to validate such object we're actually re-building all the validators (which is essentially the compile part of the process).

I've tried to move validators build logic (currently in RefValidator::ensure_validators) into the compilation process, but it get's though due to potential circular dependencies with references.
In order to reduce the code changes and keep performances in consideration I'm opting for storing the validators behind a RwLock.
The decision of using it is mostly related to the fact that respect a Mutex it does have similar performances but at the same time allows multiple readers (in our case multiple parallel validations).

Performance improvements are over well 45% improvement for validation. Raw data in here

PS. once again this was spotted while working on maci-split-validators 😎

[Stranger6667]

A minor typo :) Otherwise it is a great improvement! Great to see it

[Stranger6667]

-95.816%

Just wow :)

[macisamuele]

To be honest, in realistic cases (where references point to big schemas) I would expect even bigger improvements.
Think about Swagger specifications where the reference points to the schema of an endpoint response (ie. interacting with Kubernetes)