[xla:cpu] Remove TraceMe annotations from individual thunks

Because all thunks in XLA:CPU are non-blocking, TraceMe annotations are super confusing as they only capture the time that it takes to launch async tasks into the thread pool.

Instead ThunkExecutor should use TraceMe producers/consumers to correctly profile async thunks execution.

Also it's very easy to forget to add TraceMe to individual thunks (judging by the fact that not all thunks have TraceMe annotations), and it's better to annotate them in a single place.

PiperOrigin-RevId: 729734307