Fix broken data sourcing benchmark

[llucax]

The benchmark stopped working after multiple breaking changes accumulated over time:

• Channel type mismatch: ChannelRegistry.get_or_create() was called with ComponentMetricRequest but DataSourcingActor internally creates channels typed as Sample[Quantity].

  Fix: use Sample[Quantity].

• Missing api_client_streaming flag: MockMicrogrid defaults to not streaming, so no mock data was ever sent.

  Fix: pass api_client_streaming=True.

• Actor lifecycle: The old @actor decorator auto-started in __init__, but the new Actor class requires async-with. The benchmark sent requests before the actor was running, so subscriptions were missed.

  Fix: restructure to send requests inside the async-with block and manually start streaming tasks after the actor has subscribed.

• Cleanup failure: mock_grid.cleanup() failed because init_mock_client doesn't set mock_resampler.

  Fix: reset _CONNECTION_MANAGER directly.

• Consume tasks never stopped: Broadcast channels were never closed, so ReceiverStoppedError never fired.

  Fix: cancel consume tasks after streaming completes.

[llucax]

I guess for this to not break again (soon) we should probably run the benchmarks as part of the CI...

[llucax]

@simonvoelcker maybe we can merge this before your PR so you can actually run the benchmark before and after the changes to see how they compare?

[simonvoelcker]

@llucax good idea and thank you for fixing these issues. I had identified a few myself but was in disbelief because the benchmarks didn't even work for me in the oldest versions, right after they were introduced.

[llucax]

@llucax good idea and thank you for fixing these issues. I had identified a few myself but was in disbelief because the benchmarks didn't even work for me in the oldest versions, right after they were introduced.

Yeah, that's very weird. To be honest, I didn't investigated myself, I just delegated to AI, so I'm not sure how it actually happened that it doesn't work for old versions. But it could be wrong/lack of deps pinning.