fix(tracing): make no-op starts idempotent

[sylvesterkaczmarek]

Summary

• make repeated NoOpTrace.start() calls idempotent
• make repeated NoOpSpan.start() calls idempotent
• prevent a second mark_as_current=True start from overwriting the original ContextVar token

Normal trace and span implementations already ignore repeated starts. The no-op implementations did not, so a second start replaced the saved token with one whose previous value was the no-op object itself. A single finish(reset_current=True) then restored that object as current instead of clearing it, leaking tracing context when tracing is disabled.

Test plan

• added focused regression coverage in tests/test_noop_tracing_start_idempotent.py
• GitHub Actions

Issue number

N/A

[seratch]

I confirmed that explicitly starting the same no-op trace or span twice overwrites its saved context token and leaves the no-op object current after one finish. However, the documented lifecycle is either context-manager use or one balanced manual start()/finish() pair, and I could not find an SDK path that starts the same no-op object twice.

Please provide a concrete supported scenario that reaches this repeated-start sequence and shows the resulting context leak affecting later work. Without such a scenario, this currently protects only an unsupported lifecycle composition, so I would close this PR for now.