[aclark4life]

Fixes #9939.

• When a source pixel center lies exactly on the boundary between two output pixels (for example 2755 -> 688, or 1837 -> 306), floating point rounding could make both neighbouring BOX windows reject it, so the pixel was dropped from the result.
• Shift the BOX filter edges, and the matching window range rounding in precompute_coeffs, by a small tolerance so that boundary pixels are consistently assigned to exactly one output pixel.
• Add a test checking that every source pixel contributes to the output for several affected sizes.

[radarhere]

Be aware of the other suggested solution, #9949

[radarhere]

Closing, as #9949 would seem to be the more theoretically solid solution.