fix: fix z_sub_thr termination and improve perf

[wyfo]

Using Thread::unpark std::process::exit allows session to be closed, and callback final print to be executed.
Callback lock accounted for 2.5% of z_sub_thr perf. Using atomic and fine-grained lock remove the performance impact (even if it doesn't really change the global outcome).

[Mallets]

Any change on the example should be aligned with all Zenoh bindings.
All examples should provide the same way of providing measures.
So, should all the throughput examples be updated?

[Mallets]

Just to reiterate on the point. Any change in examples needs to be coordinated across all bindings.
This is particular important in the case of throughput examples that are made in such a way the measurement is performed in the same way in all Zenoh APIs.

I'm marking this PR as invalid for the time being till further actions are taken. Either closing the PR or providing updates to all other bindings.

[wyfo]

The throughput examples are not really aligned across bindings.
Python is the only one working as expected, i.e. printing a recap at the end. The other ones declare a drop callback which is not executed, their behavior is at least misleading.
Also Rust is the only one to use a mutex (with callback_mut). C and C++ example doesn't do that (Python don't need obviously), so their is also an example alignment issue there, as well as a more serious issue of presenting example with "buggy code" of unsynchronized callback (which can be called multiple times concurrently).

That's why I think that both C and C++ example should be fixed and aligned on this more correct new Rust version.

[Mallets]

Then we need a review of the examples. Again to reiterate on the point, any change in examples needs to be coordinated across all bindings. Modifying this example right now would also bring a difference w.r.t. historical measurements.

[wyfo]

Modifying this example right now would also bring a difference w.r.t. historical measurements.

Ok, so it's better to close this PR anyway.