Compare ambit:bounds/3 against the equivalent ambit:shape/3 call.

To avoid the ordering bias in the original benchmark (bounds always ran first, shape always second, so bounds always paid the cold-start cost for code both paths share — neighbors/1, projection, cell_corners_and_centroid, etc. — while shape ran warm), this version: 1. Warms up BOTH paths once before any measurement is taken. 2. Alternates which function is timed first on each iteration. 3. Runs N timed iterations per function and reports the median, min, and max instead of a single sample, since a single BEAM measurement is vulnerable to GC pauses and scheduler jitter.