fix(radar): decode FMI precipitation as a rate, not metres of water

[eetu]

The rain tab's forecast frames rendered essentially blank while FMI's own app
showed a clear precipitation field for the same hour.

Root cause

Precipitation1h is GRIB2 parameter 0-1-7 — precipitation rate in
kg m⁻² s⁻¹ (mm/s), not metres of water. The decoder multiplied by 1000 instead
of 3600, so every value came out 3.6× low, was rounded to 0.1 mm, and then had
anything below the ramp's 0.1 threshold dropped. Net effect: only cells above
~0.36 mm/h drew at all, and everything up to 3.6 mm/h drew in the palest blue.

It also explains why the "ei sadetta ennusteessa" hint never appeared on those
frames — frame.max was non-zero (0.72 for the reference hour), so the frame
wasn't classified as dry even though it painted almost nothing.

Verification

Checked against FMI's own point-forecast API rather than only unit tests:

• The Tampere cell decodes to 2.7778e-5; opendata.fmi.fi/timeseries reports
  0.1 mm for that hour. × 3600 = 0.1 ✔, × 1000 = 0.028 ✘. Every decoded
  value is an exact multiple of 0.1/3600.
• 10 of 12 forecast hours at the saved location now match Precipitation1h
  exactly; the two outliers differ by 0.1 mm, which is nearest-cell pick vs FMI's
  bilinear interpolation.
• 6 off-centre probes match exactly. Worth doing because Tampere sits mid-box,
  where a flipped-rows bug would be invisible — these confirm orientation too.
• The reference frame now covers ~78% of the visible map in structured
  0.1–1 mm/h precipitation, peaking at 1.4 mm/h.

For the observed/forecast ramp unification, tested through halo's own proxy: the
styled tile returns 9/9 halo ramp colours with zero foreign colours; the unstyled
tile returns FMI's palette instead (proving the SLD is what's doing the work);
and a deliberately broken SLD falls back to a tile byte-identical to the
unstyled one.

cargo test 24 + 18 green (4 new radar tests), yarn validate clean.

Also in here

• Sharper grid. The old 9.0° × 3.4° box produced 401 native cells, tripping
  MAX_GRID_SIDE and halving the grid to 201×76 — of which only ~22% of the
  width was ever on screen at the default zoom. Now 4.4° × 2.4°, shipping at
  native ~2.5 km resolution (196×107, stride 1).
• Retuned ramp to 0.1–30 mm/h; Finnish rain is mostly 0.1–5 mm/h, which the
  old 0.1–40 ramp spent entirely on blues.
• Mercator row placement. imageOverlay stretches linearly in the
  projection, but the grid is regular in latitude, so rows landed ~2.6 km too far
  north mid-box. Measured placement error drops 2.57 km → 1.26 km, the residual
  being half-cell quantisation.
• One ramp across the whole timeline. Observed frames move from
  Radar:suomi_dbz_eureffin to Radar:suomi_rr_eureffin (rain rate, band =
  mm/h × 100) and are restyled with halo's own ramp via SLD_BODY, generated
  from PRECIP_STOPS. Previously the palette jumped as you scrubbed past "now"
  and the legend described only half the timeline — dBZ can only be labelled
  qualitatively, and FMI's own style names its bins
  kohtalainen/sakea/hyvin sakea. Retuning PRECIP_STOPS later now updates both
  halves automatically.
• Proxy fallback. If a tile carrying SLD_BODY comes back as a WMS
  ServiceException (XML, sometimes with a 200), the proxy retries once without
  it, so the radar degrades to FMI's palette rather than an empty map.
• Location marker — accent ring + dot at the saved coordinates, in
  markerPane so frames never paint over it and non-interactive so it doesn't
  eat the long-press scrub gesture.

Worth knowing

• halo's forecast will still read lighter than FMI's app at short lead times.
  That view is a radar extrapolation nowcast; this is the HARMONIE model, and
  HARMONIE genuinely forecast only 0.1–0.4 mm over Tampere for the hour in
  question — FMI's human-edited pal_skandinavia agrees. FMI's open WMS
  publishes no nowcast layer (all 116 Radar:* layers are observed products), so
  there is nothing to switch to. That gap is meteorology, not code.
• rr shows slightly less fine texture than dBZ. Reflectivity picks up faint
  echoes the rain-rate product doesn't, and halo's ramp starts at 0.1 mm/h where
  FMI's rr style starts at 0.07 — so the very lightest returns now read as dry.
  That's the cost of both halves being the same quantity.
• Observed frames are slightly lighter than before, since RADAR_OPACITY
  (0.75) and FORECAST_OPACITY (0.6) collapsed into one OVERLAY_OPACITY;
  differing alphas would have re-introduced a jump at "now".
• The SLD depends on FMI keeping dynamic styling enabled. The fallback covers
  them turning it off — if the observed frames ever show FMI's louder cyan/green
  palette, that's it firing (retrying with FMI's own style in the log).