feat: server-side time-series downsampling + dashboard/API cleanup

[simontaurus]

Summary

Downsample large time series on the server (pgstack downsample_tool_channel RPC) and transport only what the plot can show, driven from the dashboard zoom. Plus a set of dashboard and API cleanups.

Downsampling

• DownsampleParams (sample / average / minmax, max_points, bin_size, edge_anchors) on the read/load params.
• PostgrestDatabaseDriver routes to the RPC with graceful fallback (missing RPC / older stack / SQLite all still work); LocalDatabaseDriver stride-decimates.
• View: per-channel method resolution, cache bypass while downsampling, and a "Load current range" button that re-fetches the box-zoomed window at finer resolution.

Dashboard fixes

• Link plot x-ranges via one shared DataRange1d created at figure build (a deferred reassignment raced the render and left some plots unlinked).
• Bridge the Reset event through a ColumnDataSource + CustomJS (figure.on_event(Reset) does not reach the server through Panel).
• Fit the Wunderbaum tree to the sidebar (was overflowing to a fixed 800px).

API / cleanup

• auto_fetch now respects the end picker (no longer pins the end to now); the live view keeps its own now-window and is unaffected.
• High-level bulk store store_channel_data_bulk + a shared seed helper (_demo_data) used by demo/bench/tests; consolidated the Playwright screenshot helpers into docs/_screenshot_utils.py.
• Config models use plain pydantic BaseModel; the param classes are re-exported from the user-facing module so callers need not import the mixin.
• Consolidated read_tool_channel_raw into the concrete TSDCMixin (no v1/v2 override duplication).

Tests / assets

• Unit tests + gated pgstack integration tests + a benchmark; the downsampling demo; regenerated screenshots and README section.

Verified: offline suite (incl. new bulk-store + config round-trip), 8 integration tests against a live pgstack, and a demo smoke run.

Note: the .content scroll relies on opensemanticworld/panelini#37.