Load detectors separately and handle multiple detector banks

[SimonHeybrock]

Loki (and many other instruments) will have multiple detector banks. We need to handle this in a convenient manner. Eventually this will likely also require input from IDS/IS (at least in certain cases), but I believe our current understanding is sufficient to proceed:

• Multiple banks (NXdetector groups in a Nexus file) will be handled using a param table in Sciline.
  • Note: implementation details may change regarding Sciline and param tables, but this is mostly orthogonal to what we need to do here, i.e., we can proceed).
• Split loading of monitors and detector data #82 (comment) mentions loading of non-monitor and non-detector data. This will be required in the medium-term future (sample logs, chopper logs, ...), but currently we are not using any of that information. That is, we can essentially ignore this aspect for now, since it can easily be added later.
• Gather results from all banks once we are in $(Q, \lambda)$ or $Q$, but before normalization, sum (concat if events) them.
  • For now this means inserting an extra step in the workflow, see Explicit merge / accumulation strategies and customization points #97. This may be away of Sciline's way of handling param tables is changed. Do not worry about this for now, just add the step.

Complications, to ignore for now:

• Beam center. See SANS beam center correction may need to be applied differently #28. Does it apply only to the rear bank in Loki? Are the others fixed enough, with well-known positions.
• Will we have a different direct-beam for the other banks? This is probably easy to implement, by including direct beam files in the param table, if required

Summary of work to be performed:

• Implement Load monitors separately from the remainder of the files #99 first.
• Instead of loading an entire file, load an individual NXdetector.
  • You can assume that all detectors are direct children of NXentry/NXinstrument. The detector names are defined in the workflow, either as a single param DetectorName (when processing a single bank), or as a sciline.ParamTable (or using https://scipp.github.io/sciline/generated/classes/sciline.Pipeline.html#sciline.Pipeline.set_param_series).
  • You can assume that position information is self-contained, i.e., there are no depends_on chains leading outside a detector. We can thus directly use snx.compute_positions on the individual detector.
  • Add a filter to the existing load function such that no detectors are loaded there. This is essential to avoid potential massive overhead from large event-mode detectors.
• The functions for "extracting" detectors for the loaded data probably need to preserved for supporting legacy NeXus files (ISIS). They should be moved to that submodule if feasible.
• Add a step for merging intermediate results before performing the normalization. This will require merging the numerator as well as the denominator (the denominator depends on the detector bank via the solid-angle factor). Tracked in Explicit merge / accumulation strategies and customization points #97.