runtime: RSS regression in gopher-lua benchmark likely at https://go.dev/cl/658036

[mknyszek]

https://perf.golang.org/dashboard/?benchmark=GopherLuaKNucleotide-16&unit=peak-RSS-bytes&repository=go&branch=master&platform=linux%2Famd64&days=30&end=2025-05-29T20%3A45

I suspect this might be related to the fact that ptrBuf is created on each P, but isn't actually used if the green tea experiment is off. Or it's just something related to some new fixed-size overhead. We only see this regression in gopher-lua, which is one of the few small-heap benchmarks we track RSS for.

[gopherbot]

Change https://go.dev/cl/678535 mentions this issue: runtime: reduce per-P memory footprint when greenteagc is disabled