Add clear_streams API for cleanup before exit

[zcbenz]

Windows horror story: we should not do anything complicated during thread/process teardown (i.e. in the callbacks registered by system atexit explicitly via APIs or implicitly via C++ static/thread_local keywords), joining threads would cause deadlocks, and calling any CUDA API would fail.

The solution proposed here is to just provide an API to let users invoke before thread/process exit:

    std::thread t([]() {
      auto s = new_stream(default_device());
      eval(arange(10, s));
      clear_streams();
    });
    t.join();

The API call is optional, if users don't use it the program would still work normally on macOS/Linux that everything gets synchronized and cleared implicitly by us, except for Windows where the program would crash on exit, which is actually fine for most cases.

[angeloskath]

👍