boolmesh: scratch bound for batch paths (parallel-batch stealing, union_many intermediates)

[GeneralPawz]

Left open by #226, which made boolmesh's declared scratch an upper bound for single booleans: `ScratchRequirement::Affine` = base + per-triangle × n + per-worker × workers. Two batch cases aren't covered by measurement yet:

• Parallel batch: with both `parallel` and `parallel-batch`, a worker waiting in a join can steal another pair's boolean, so more booleans can be in flight than there are workers. The per-worker term (base + 16 KiB) is reasoned, not measured.
• `union_many` intermediates: intermediate results can have more triangles than the original inputs, so a bound in input triangles may not hold.

Also: a one-thread `CpuExecution` installs no pool, so the provider runs on the ambient pool. #226's provider-side re-check catches this, but only when a budget is set.

Done when

• Batch booleans and `union_many` are measured under forced worst-case schedules, as in boolmesh scratch_bound flakes under load: 24-triangle SymmetricDifference peaks above the declared 4096 B/triangle #226's `Schedule::ColdPool`.
• The declaration covers them, or those paths refuse budgets they can't bound.
• The one-thread case runs on a pool of its stated width.