Windowed live-data fetching + coordinated ECharts flush timers

[Bilchreis]

Summary

Live plots previously fetched a parameter's entire value history on every load, then relied on the client to trim it down to the visible time window wasteful for parameters with long histories. This PR makes live-mode fetches window-aware end to end, and fixes a related client-side performance issue in the chart flush loop that was scaling badly with plot count.

Backend: window-aware live data fetching

• Builders.fetch_param_data/2 now takes a mode. For :live, it fetches only the configured default time window (fetch_param_data_in_range/3) instead of full history; :historical keeps the old unbounded fetch.
• Added Builders.any_data?/1 to distinguish "no data in the live window" from "no data ever recorded" without re-fetching full history, and Builders.plot_available?/3 to use that distinction when deciding whether a plot can render (a parameter with only old data outside the live window is still "available", just showing an empty chart).
• If the live window is empty and the parameter has no data at all, we fall back to reading from the device (read_from_device_if_empty/2), preserving existing bootstrap behavior.
• Range buttons and the default active range index are now configurable via config :secant_service, :plot_db (falls back to the previous hardcoded defaults), and Builders.default_live_range_ms/0 centralizes resolving the default window against the current time so the fetch path and each dtype's initial x-axis window can never disagree.
• All plot spec modules (Calibratable, Drivable, Readable, PlotDB) and dtype option builders (Scalar, Enum, Struct, ArrayHeatmap) updated to thread mode through and use the shared helpers instead of duplicating range-computation logic.

Frontend: coordinate ECharts flush timers

Each mounted chart ran its own setInterval doing a full series rebuild + setOption every tick, regardless of whether new data had arrived, so CPU cost scaled with timer frequency × plot count. This PR:

• Replaces per-chart timers with a single shared ChartFlushDriver ticking all mounted charts once a second.
• Adds a _dataDirty flag so the expensive part of a flush (trimming, series rebuild, heatmap visualMap recompute) only runs when data has actually changed since the last flush.
• Adds a shared IntersectionObserver (ChartVisibilityObserver) so fully off-screen charts skip setOption entirely; buffered data isn't lost, it's just not rendered until the chart is visible again.
• The x-axis window still slides every tick regardless of dirty/visibility state, so live charts keep visibly advancing with wall-clock time.
• Generalized point-timestamp access (_pointTs) and range-expansion backfill (_earliestBufferedTs) to handle both plain [ts, ...] points and enum's {value: [ts, y]} point shape, and to correctly trigger a backfill fetch even when a series has zero buffered points.