perf: reuse triangulation/point-location in grid interpolation (drop Dask)

[JanStreffing]

Problem

scipy.interpolate.griddata rebuilds the Delaunay triangulation of the source points and re-runs the point-location of the (fixed) target points on every call. Interpolating the ~60 vertical levels of the 3D CO2 field — and the per-field 2D emissions/albedo loops — repeats that identical work N times. The 3D CO2 step then wrapped it in a dask.distributed LocalCluster whose process startup + a pathological ~150 s worker teardown dwarfed the few seconds of actual interpolation.

Profiling (TCO95/feomdyn) showed the CO2 interpolation dominated by exactly this: LinearNDInterpolator.__call__ re-doing find_simplex at ~2.7 s × 60 levels.

Fix

New ocp_tool/interp_utils.ReusableGridInterp: builds the Delaunay + KD-tree once, and the target-point barycentric weights once — so each subsequent field is just a gather + weighted sum. Applied in:

• co2_interpolation — horizontal loop (was per-level griddata + Dask). The vertical interp1d per-point loop is collapsed to a single vectorised interp1d(axis=0) (after the horizontal nearest-fill there are no NaNs, so every column shares the full level set). Dask removed entirely.
• field_interpolation — the 2D CO2-emissions (linear, fill_value=0) and albedo (nearest) per-field loops reuse one triangulation / KD-tree across fields.

Correctness

Bit-identical to griddata to interpolation precision (the barycentric weights are the same ones LinearNDInterpolator uses; max abs diff ~1e-19), and bit-for-bit after GRIB packing — a full-pipeline A/B (old Dask path vs new) gives grib_compare: IDENTICAL.

Impact (TCO95/feomdyn)

• CO2 interpolation: ~101 s → ~6.7 s (~15×)
• Full run_ocp_tool.py: 211.6 s → 138.6 s
• No dask/distributed dependency in the interpolation path.
• The win grows with resolution: the triangulation build, previously repeated per level/field, is amortised once.