contour: remove redundant np.empty_like allocation in coordinate transform

[Melissari1997]

Reduce peak memory in contours() by removing a redundant np.empty_like allocation during the coordinate-space transform step.

np.interp already returns a new array, so the intermediate np.empty_like allocation was immediately overwritten. Replaced with np.column_stack which builds the output directly from the two interpolation results.

Closes #3351

Backend coverage: numpy only (the transform runs after backend dispatch completes, on the merged numpy results common to all paths).

[Melissari1997]

PR Review: contour: remove redundant np.empty_like allocation in coordinate transform

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

None.

What looks good

• Correctness is preserved. coords is a freshly-allocated array returned by _stitch_segments (np.column_stack at contour.py:299). No other reference aliases it — _deduplicate_lines only reads the array and places the same object into the deduped list, so in-place mutation of coords in the transform loop is safe.
• Memory improvement is real. The old code allocated an extra np.empty_like(coords) per polyline that was immediately overwritten. The new code avoids that allocation, reducing peak memory by one full coordinate array.
• Surgical change. Only 3 lines changed. No new imports, no new dependencies, no refactoring of surrounding code.
• Backend coverage is correctly scoped. The transform runs after backend dispatch completes, on the merged numpy results common to all paths.
• Existing tests cover the changed code path. The test suite exercises the numpy contour path with several test cases that all hit the modified transform loop.
• Edge cases are handled correctly. Empty results, single-point polylines (already dropped by _stitch_segments), and closed rings all behave identically to the old code.

Checklist

• Algorithm matches reference/paper
• All implemented backends produce consistent results
• NaN handling is correct
• Edge cases are covered by tests
• Dask chunk boundaries handled correctly
• No premature materialization or unnecessary copies
• Benchmark exists or is not needed
• README feature matrix updated (if applicable)
• Docstrings present and accurate