Speedup Local-Non-Local Context's data operations

[franz1981]

This is addressing 3 performance problems:

• local/non-local context data's operations shouldn't use costy synchronized on existing maps (now that biased locking is gone)
• local/non-local context data's lookup/remove doesn't really require to create any map, if not already present
• local-non-local duplicated ctx data operations are not piggybacking on delegated ones, forcing using default subptimal implement ones, instead

This PR is pushing improvement furthers, but leveraging the facts that:

• default's ConcurentHasMap constructor doesn't initialize the internal table, making very cheap to allocate one
• most of the time there's no concurrent data allocations, which means we can save using synchronized and just use compareAndSet saving another strong memory barrier (in x86, but likely the same in ARM), speeding up common use cases

[vietj]

before doing anything here, we should first validate and finish this PR #5048 which has an impact on the overall design, since I would like to back port this PR to 4.x

[franz1981]

Let's wait till you sync with @cescoffier about the new API and the time to backport it: this one is pretty small and contained, so, depending for how long we will keep on using the current API, could be a waste or a good idea to bring it in.

[franz1981]

@vietj if merged I have to fix https://github.com/quarkusio/quarkus/blob/4e6908f522190127a28af7c96806c06cc0b0a60e/extensions/vertx/runtime/src/main/java/io/quarkus/vertx/core/runtime/VertxLocalsHelper.java#L25 which is wrongly allocating a local map, although not really required.