Scale Ri CAT thresholds by actual layer thickness (Δz) instead of the altitude ramp

[roznet]

Follow-up to #533 / PR #534 (see designs/meteorology-decisions.md §25).

The NWP positive-Ri bias that motivated loosening the Miles-Howard tiers is a function of the layer thickness the Ri was computed over, not of altitude per se. The current _ri_threshold_scale ramp (×1.0 at ≤10,000 ft → ×2.0 at ≥20,000 ft) is an altitude proxy that:

• applies the same multiplier to a GFS 25 hPa layer and a UKMO 50 hPa layer at the same altitude (raised by the review bot in rounds 2 and 5 — UKMO/Météo-France were never spot-checked);
• was calibrated against ICON/GFS/ECMWF on a single route.

Since the #534 geometry fix, every CAT layer carries its real physical extent (base level → flagged level), so Δz is directly available per Ri value. Scaling thresholds by Δz / Δz_ref (capped, e.g. at the current ×2):

• subsumes the altitude ramp (thickness grows with altitude automatically);
• self-corrects per model with no per-model configuration;
• removes the residual half-layer bias of evaluating the scale at the layer midpoint.

Bulk-Richardson literature supports depth-dependent critical values. Calibrate Δz_ref/cap so ICON/GFS/ECMWF behavior on the EGNY→EGKB pack and an FL300 jet-flank case are preserved, then spot-check UKMO/Météo-France.