[RhysSullivan]

Tests the fix in #14.

The new workerd.bounds test fills an 8 MiB store until it first evicts. It then checks two things:

• the store is at or below the 90% low-water mark;
• twenty idle exports of one 10-span trace each, about half a MiB in all, are stored without evicting any trace the full store held.

It also expects exactly one eviction pass in total.

On main (45c0e60), the same 20 exports evict 5 of the stored traces. The first assertion fails as well: the store sits at 8,376,320 bytes, at the bound.

No restart test: whether one pass finishes a drain depends on how many pages each eviction frees, so a deterministic one would need a contrived pass budget. The restart case was verified in a 1 GiB measurement instead (see #14): the drain reached 922 MiB after workerd unloaded the collector mid-drain.