Issue #3620435: Say on the performance page that a kernel test cannot see cache work

The measuring rules on the performance page told a reader to warm the caches, name the scenario and say whether a figure was measured or inferred. They did not mention the one thing that makes a cache measurement meaningless: KernelTestBase puts every cache bin on the memory backend, so a cache read and an invalidation are not database statements there. A kernel budget test reports a change in cache traffic as costing nothing, and reports it as a pass.

The guidance was written for #3616990 and lives only on !246, which is parked behind a postponed issue and eighty-two commits behind, so nobody reading the performance page would find it. It stands on its own: it is true whether or not that change ever lands.

The sentence above the list counts the rules, so it says four now rather than three.

Documentation only, no code touched.