mask netcdf output by wet-cell topography instead of value==0

[JanStreffing]

PR #799 set any accumulated output value below 1e-30 to _FillValue. That also masks legitimate ocean zeros: ice-free sea-ice concentration (a_ice and all other 2D ice variables) and vanishing 3D fields such as IDEMIX energy were written as NaN instead of 0, contrary to the CMIP/CMOR convention. The masking was applied uniformly to 2D and 3D fields.

Replace the value-based test with a proper wet-cell mask derived from the mesh level arrays (ulevels/nlevels for elements, ulevels_nod2D/ nlevels_nod2D for nodes). Only cells above a cavity or below the bottom topography are written as _FillValue; every wet cell keeps its averaged value, including a genuine zero. 2D fields and non-spatial vertical axes (density / ice classes) are never masked.

@chrisdane Can you check if this fixes what you flagged in #799 ?

[JanStreffing]

Adding a data point in favor of this fix: #939 is another symptom of the value-based masking from #799, this time on river runoff. Same pattern as #821: a sparse 2D field where _FillValue lands at every node that should hold a legitimate zero, so xr.mean() over nodes divides by the nonzero count and inflates the apparent magnitude by N_total / N_nonzero ≈ 2.6×. Worth prioritizing for 2.7.1.