Aggregate lake pieces per cell in lake_from_gdf, like RIV/DRN celldata aggregation

[bdestombe]

What

lake_from_gdf now accepts per-piece input — multiple rows per cellid for one lake, as produced directly by nlmod.grid.gdf_to_grid — and combines the pieces into one lake-GWF connection per cell with an area-weighted bed resistance. It also tolerates float-level noise in numeric per-lake settings.

Why

Two defects when the input is not pre-aggregated by the caller:

1. Per-piece rows multiply-count the lake-aquifer exchange. MODFLOW 6 applies each VERTICAL lake-GWF connection over the full cell area, and lake_from_gdf turns every row into its own connection with bedleak = 1/clake. A cell holding two pieces of one lake (6000 m² at clake=10 and 2000 m² at clake=20 in a 10000 m² cell) gets two full-cell-area connections totalling 0.15 * cell_area of conductance, where the pieces physically justify (6000/10 + 2000/20) / 10000 = 0.07 * cell_area.
2. Exact-equality single-value check rejects aggregated inputs. _get_and_check_single_value compares strt (and outlet settings) with ==. Area-weighted per-cell stages carry float-level noise (a 1e-12 difference), which raised AssertionError: A single lake should have a single strt for no physical reason.

RIV and DRN already have exactly this aggregation step in nlmod: nlmod.gwf.surface_water.aggregate collapses per-piece celldata to one reach per cell with area-weighted parameters before build_spd. The LAK path lacked the equivalent, forcing every caller to hand-roll it. (Encountered wiring the NHFLO Bergen ponds — NHFLO/models#127, NHFLO/tools#58 — where the workaround currently lives downstream.)

How

• New _aggregate_connections_per_cell(lake_gdf, ds), invoked per lake when the cellid index has duplicates: combined clake = cell_area / sum(piece_area / clake), so the VERTICAL connection conductance cell_area / clake equals the summed piece conductance; other columns take the first piece's value (they are per-lake settings, still validated by _get_and_check_single_value). Raises a clear ValueError when pieces come without geometry, since the areas are then unknown.
• _get_and_check_single_value uses np.allclose(rtol=1e-8, atol=0) for numeric columns; exact comparison is kept for strings (boundnames, timeseries names, lakeout).
• nlakeconn consequently counts cells, not pieces.

Behaviour for already-aggregated input (one row per cell per lake) is unchanged.

Tests

Two regression tests in test_013_surface_water.py, both shown to fail on unchanged dev before the fix:

• test_lake_from_gdf_aggregates_pieces_per_cell — fails on dev with 3 connections at bedleak=0.1/0.05 instead of 2 connections at the area-weighted 0.07/0.1; also pins nlakeconn == 2.
• test_lake_from_gdf_accepts_floating_point_strt_noise — fails on dev with the AssertionError on a 1e-12 strt difference.

test_gdf_lake (existing) still passes; ruff check/ruff format add no findings relative to dev.