fix(Redis): use shared_ptr for NotificationCenter, harden AsyncNotificationCenter::stop()

[matejk]

Summary

• Use shared_ptr internally in Redis::Client for the AsyncNotificationCenter. When std::atomic<shared_ptr> is available (C++20 __cpp_lib_atomic_shared_ptr), use it directly; otherwise protect the shared_ptr with a std::mutex. This replaces the raw-pointer-with-no-op-deleter approach that could leave callers with dangling pointers if the Client was destroyed first.
• Harden AsyncNotificationCenter::stop() to unconditionally wait for and join the dequeue thread, preventing potential issues if stop() is called after the thread has already exited.
• Add missing connection guard to RedisTest::testPool.
• Enable Redis tests in windows-2025-msvc-cmake CI job.

Test plan

• Foundation AsyncNotificationCenterTest passes (14 tests)
• Redis library and test runner build cleanly
• Verify on Windows CI (Redis tests should skip gracefully without a server)

[aleks-f]

Client may delete the pointer, leaving holders with a false ownership of a dangling pointer. It was a quick and dirty temporary hack until c++20 and atomic shared_ptr becomes available everywhere, but not a good idea really.

Proper solution is to use shared_ptr internally, protect it with mutex when atomic shared_ptr is not available.

[matejk]

Client may delete the pointer, leaving holders with a false ownership of a dangling pointer. It was a quick and dirty temporary hack until c++20 and atomic shared_ptr becomes available everywhere, but not a good idea really.

Proper solution is to use shared_ptr internally, protect it with mutex when atomic shared_ptr is not available.

Applied requested changes.