Performance regressions in 15.0-RELEASE

[dag-erling]

Summarizing from memory, there are reports of performance regressions in 15.0 in the following areas:

1. jemalloc 5.3.0 is allegedly slower than its predecessor, though apparently not for all use cases, and one person reports that they were previously suffering from runaway heap fragmentation which jemalloc 5.3.0 fixed, so we don't want to rush to revert.
2. clang takes longer to start due libllvm and libclang now being dynamic, which greatly impacts poudriere build times (it's probably particularly bad for configure scripts and for C codebases with many small source files, while C++ programs are probably less affected because startup time is proportionally less important); this patch should help by adding an option to turn libllvm and libclang back into internal libraries so they're linked at build time rather than load time.
3. namei lookups, or to be precise, namei lookup failures, take longer
4. system call overhead is allegedly up.
5. execve() is allegedly slower than in 14.3, by more than can be explained just by the previous item; this is also going to noticeably impact both world and poudriere build times.

We should make sure these reports are investigated and any actual regressions are addressed, but we should also think about what we can do to catch such regressions earlier in the future.

[dag-erling]

Regarding jemalloc, we also raised LG_VADDR from 48 to 64 due to LA57. This is expected to have a minor impact on performance, which we should keep in mind when benchmarking.

[markjdb]

Mark will see if we can run will-it-scale as part of stabweek, and keep track of results over time.