Approximate sizes of stack and visisted when collecting accumulated values

[MichaReiser]

This PR uses the input-output length to approximate the number of items added to stack and reserves stack.len() additional items in visited to avoid costly resizes. This reduces the performance regression in astral-sh/ruff#14760 from 10.3% to 9%, which seems worthwhile.

The downside is that it can result in a too large stack if a query has many outputs but only few outputs or in a too large visited set if many inputs are overlapping. I think this is acceptable, considering that the data structures are short-lived.

Performance

This shows a 9% performance improvement for the accumulator benchmark.

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	5314d4c
🔍 Latest deploy log	https://app.netlify.com/sites/salsa-rs/deploys/67502214ff59100008541079

[codspeed-hq]

CodSpeed Performance Report

Merging #622 will not alter performance

_{Comparing MichaReiser:small-accumulator-perf (5314d4c) with master (e68679b)}

Summary

✅ 9 untouched benchmarks