Refactor AdePT into 2 libraries 4: Fix transport lifetime ownership and shutdown

[SeverinDiederichs]

This PR cleans up the lifetime of the shared AdePTTransport:

Instead of keeping the transport alive through a static owning shared_ptr, the static storage now only keeps a weak_ptr, while the real ownership remains with the AdePTTrackingManager instances.

This avoids destroying the transport very late during process teardown, where CUDA stream cleanup can already be in an invalid shutdown state.

Another explicit guard was added in the AsyncAdePTTransport:
In this wait loop for work of the G4 workers:

    while (gpuState.runTransport && std::none_of(eventStates.begin(), eventStates.end(), needTransport)) {
      using namespace std::chrono_literals;
      std::this_thread::sleep_for(10ms);
    }

The GPU steering thread can enter the sleep at shutdown, because the needTransport is already false, but the runTransport is not yet switched to false in the AsyncAdePTDestructor. If it is then switched to false by the worker in the shutdown, then one more iteration is run, while CUDA might already be in teardown, leading to a crash. This is prevented by a simple guard.

This PR is needed for upcoming restructuring, which exposed this life-time-management problem

It was verified that this PR

• Changes physics results
• Does not change physics results

[phsft-bot]

Can one of the admins verify this patch?

[agheata]

Looks good, LGTM