Export buffered dashboard data to NeXus files

[SimonHeybrock]

Closes #1351.

Scientists need data that is never written as a run, e.g. alignment scans, to reference later. This adds an Export data button to the dashboard header. It opens a wizard modeled on the plot wizard:

1. Select workflow and output (the plot wizard's first step).
2. Select the export type. Time series is the buffered history of a 0-D time-series output, optionally correlated with up to two other time series. It is offered only for 0-D outputs. Current value uses the plots' window modes: since run start, latest update, or a sum/mean over a window.
3. Select sources (the generic configuration form).
4. Preview what is buffered (groups, shapes, time ranges) and download a NeXus file.

Only what DataService holds can be exported. Previously a key kept history only while a plot subscribed to it with a history extractor, so a scan done without the right plot open left nothing to export. A permanent TimeseriesRetention subscriber now keeps history for every 0-D time-series output, within the existing 20 MB per-key TemporalBuffer cap. A window aggregation is served only if the buffered updates reach back over the whole window, and is otherwise refused with what is buffered. 0-D time series are always covered. Other outputs are covered only while a plot in window mode buffers them, and then only for that plot's window.

Correlation. The file holds the unbinned per-update table, not the histogram: each data point carries the axis value in effect at its time as a coord, and each axis is also exported as its own time series. The lookup is shared with CorrelationHistogramPlotter.

NeXus. scippnexus cannot write DataArrays as NXdata, so nexus_writer.py writes one NXdata group per source with h5py. Tests check that scippnexus loads every group back to the exported DataArray; files also open in nexusformat (NeXpy). Source names with / (common for f144 PVs) are written with _, and the original names are kept as group attributes. The format is a field in the export params, so more formats can be added.

Transfer. pn.widgets.FileDownload sends the file as base64 over the websocket, which runs on the IOLoop all sessions share. Exports are therefore capped at 50 MB.

Placement of the modal. The modal's holder is passed into PlotGridTabs as an overlay next to its own modal container. In the header, panes the wizard builds after load stayed invisible (#1154). Wrapping the main content to hold it broke plot geometry (caught by render_geometry_test).

Open question: retention memory

Keeping history for all time series is not free. Instruments have 87 (DREAM) to 167 (ODIN) time-series keys. At 1 Hz a key fills its 20 MB cap in about a week, so a long-running dashboard could grow to 1.7–3.3 GB. Only keys with running workflows get data, so the real figure is lower. A smaller cap for retention-only keys, or a global budget, belongs with #1274. Should that land before this, or is it a follow-up?

Test plan

• New unit tests cover the writer round-trip, export collection (retention, UTC times, correlation, name collisions, / names) and the wizard steps.
• A Playwright run against the fake backend downloaded a correlated monitor/motion export and a detector image. Both load with scippnexus.
• Export a correlation of a detector or monitor total against a motor on a real instrument, and load the file in the analysis tool of choice.
• Check the header button and the wizard with basic auth enabled (next to the logout link).

🤖 Generated with Claude Code