Revert "Remove unnecessary synchronization when creating Resource instances

[geoand]

Turns out that #34339 is worse for performance...

Verified by @franz1981.

The reason is that the creation of the BeanInstance costs more than the volatile read

[franz1981]

BeanInstance costs more than the volatile read

Yep, the relevant profile data show this stack trace (that wasn't present prior the mentioned changes):

The violet bar is io/quarkus/resteasy/reactive/common/runtime/ArcBeanFactory.createInstance

And the cost of such is (for Techempower plaintext, which is heavily CPU bound) is ~9% of the cpu cost for io/quarkus/resteasy/reactive/server/runtime/QuarkusResteasyReactiveRequestContext.invokeHandler, and given that the test is CPU bound, it directly translate in a ~9% max throughput drop (from 4.5 M req/sec to 4.1 M req/sec).
Other scenarios (not CPU limited) would likely experience a minor drop, likely, but it still seems a work we can save.

[geoand]

Just adding for posterity on why this is faster:

Each JAX-RS method will be handled by its own InstanceHandler which in turn creates the CDI bean (which is a singleton by default) that includes the aforementioned method. It follows that we don't need to keep calling Arc for each request to get that instance

[franz1981]

Given that the synchronized block happen just once per endpoint, it's not that important to improve it (at a first analysis)