debug what happened with 14031

### Description

PR #14031 makes the float vector functions inlinable, but caused a drop in nightly benchmark.
We reverted the PR in #14041 , but we should try to figure it out, as we don't want unpredictable performance.

Maybe the problem is specific to AMD and we should test that. Also it would be great if we could safely add some debugging to the nightly benchmark to understand what happens with the compilation (e.g. `-XX:+PrintCompilation` or `-XX:+LogCompilation`).

Start by comparing PR #14031 with its revert in PR #14041 and reviewing the nightly benchmark results that showed the regression. Test whether the effect is specific to AMD and investigate compilation with -XX:+PrintCompilation or -XX:+LogCompilation. Done means identifying the cause and establishing a safe way to collect the needed benchmark diagnostics.

Written by the indexing model from the issue text.