Bound a cross-process shared cache's gain before building it

Before you build a cache shared by N concurrent reader processes, you can bound its gain cheaply in a single process. (1) Run the same request sequence twice in one process with the per-process cache on and any higher-level result cache off. Pass2 CPU minus pass1 CPU is the most a shared cache could save each extra reader. (2) Run with the cache size set to 0. The PSS difference (from /proc/<pid>/smaps_rollup) is the memory it could deduplicate. Compare sharing against the cache-off baseline, not against today's default size. In one case the per-process decoded-block cache held about 70% of the worker's PSS but saved only about 7% CPU, so shrinking it beat sharing it. Also check that pass 2 really exercises the cache under test: a result or tile cache in front of it can make pass 2 nearly free and hide the effect.