Experiment with single-writer boxcar::Vec

[ibraheemdev]

Testing to see if we can avoid the performance hit in #688 by using a single-writer optimized push method (which I added to boxcar in a branch). It looks like we can only do this for the ingredient_vec in Zalsa, not the page table or viewcaster vector.

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	9c3c15d
🔍 Latest deploy log	https://app.netlify.com/sites/salsa-rs/deploys/67ae89ab589dc50008d5f7fe

[codspeed-hq]

CodSpeed Performance Report

Merging #690 will not alter performance

_{Comparing ibraheemdev:switch-to-boxcar (9c3c15d) with master (03197cb)}

Summary

✅ 9 untouched benchmarks

[ibraheemdev]

Looks like this doesn't really help much. I would guess where append-only-vec is winning is avoiding per-entry initialization flags because it serializes insertion order using the spinlock, which gives it better cache-locality. That would require a completely new type in boxcar for the single-writer case, not just a new method. Either that, or the ingredients vector isn't actually where the regression is, and append-only-vec just looks better in the other cases because of the spin-lock.