Faster accumulators

[MichaReiser]

This PR implements faster accumulators.

Accumulators are relatively costly today because collecting the accumulated values requires iterating over the queries' entire dependency tree.
This cost is negligible with a cold DB but adds up for very deep trees.

For example, adding accumulators to Ruff resulted in a 30% perf regression astral-sh/ruff#14116

This PR adds a new field to AccumulatedMap that tracks whether any dependency of the current query has any accumulated values. It checks that field when collecting the accumulated values, allowing it to skip entire subtrees.

This approach should make accumulated values nearly "free" in the common cases where there are no or only very few diagnostics.

Test Plan

I added a benchmark and I verified that the perf regression in the linked Ruff PR is mostly gone after using this salsa version.

Running the benchmark locally gives me

     Running benches\accumulator.rs (target\release\deps\accumulator-51d27e7934278de6.exe)
accumulator             time:   [666.27 µs 676.29 µs 688.11 µs]
                        change: [-32.104% -30.905% -29.682%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 12 outliers among 100 measurements (12.00%)
  1 (1.00%) high mild
  11 (11.00%) high severe

where the baseline is main

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	76079d2
🔍 Latest deploy log	https://app.netlify.com/sites/salsa-rs/deploys/674f18c81e639b0008aa6168

[MichaReiser]

I don't like the naming at all but struggled to come up with something better.

[codspeed-hq]

CodSpeed Performance Report

Merging #615 will not alter performance

_{Comparing MichaReiser:micha/faster-accumulators (76079d2) with master (b7331d9)}

Summary

✅ 8 untouched benchmarks
🆕 1 new benchmarks

Benchmarks breakdown

	Benchmark	master	MichaReiser:micha/faster-accumulators	Change
🆕	accumulator	N/A	5.1 ms	N/A

[davidbarsky]

approving modulo one nit. I'll let you decide whether to merge or not.