WindowFunctionExecRange handles all range-based windows the same way: each step, it finds the window's endpoints. For a left-unbounded window like [unbounded, +3], we can make a couple improvements:

• Instead of a WindowFunctionState, we can use a cheaper AccumulatorState.
• We can release documents sooner. Policy::kEndpoints assumes the Exec needs the left endpoint, but for a left-unbounded window we only need the right endpoint.