[bos]

Fixes #992.

threshold_total called thresh on the count before and after each update, also when that count was zero. distinct_total gives it |_,_| 1, so when a key's count returned to zero it emitted 1 - 1 = 0 instead of -1. Whether a zero count reached thresh at all depended on whether the trace had merged away the updates that cancelled.

threshold_total now gives zero for a zero count and does not call thresh on it. threshold documents the same behaviour, and distinct_total keeps its closure.

Test plan

• New tests/threshold.rs: distinct_total and threshold_total (with a thresh that does not map zero to zero) on a key that leaves and returns, with all updates in one batch and with one batch for each time. Both tests fail before the change.
• A randomized comparison against a reference: 5,000 inputs in each scheduling order, including the order from reduce panics with "upper straddles batch" when its input frontier is ahead of its trace #990 (the operator runs before the arrange that feeds it), with no failures. count_total passes the same check without changes.
• cargo test --workspace.