Split load of detector/monitor from event-data load

[SimonHeybrock]

This addition is related to the new approach discussed in scipp/beamlime#215. It depends on a release of scipp/essreduce#53, i.e., CI will fail for now.

Edit: converted to draft, since I found one more dependency on event data that should not be there (for solid angle). -> This is fixed by the follow-up #155 (now merged into this branch). -> Fixes #154.
Also closes #39, closes #62.

By loading NXevent_data separately from the containing NXdetector or NXmonitor we increase flexibility, reduce overhead, and continue paving the path towards chunked processing of large files.

Before:

After:

Notes:

• By getting the detector IDs from the RawData[SampleRun], which is now the "empty" detector with no events, masks can be loaded and processed without loading potentially huge amounts of events. Even if the events will be loaded anyway this will improve parallelism, and will avoid reprocessing of masks if the event data changes.
• Chunking loading of the event data (using Sciline's map-reduce) with the "after" graph will only load new events, which is also what we need to do for live-reduction. Other detector information as well as masks can be reused. In short, the graph is "wider", which is better.

[SimonHeybrock]

I believe this was reviewed as part of #155.