Check math.round perf

Here's a surprising claim that math.round(x) is measurably slower than math.floor(x + 0.5)
https://github.com/beyond-all-reason/Beyond-All-Reason/pull/5246#discussion_r2135422692

Figure out what's going on, how bad it is and if we can do anything about it.

Start with the linked pull-request discussion and reproduce the reported comparison between math.round(x) and math.floor(x + 0.5) using a representative benchmark. Quantify whether the difference is reproducible and significant, then document the cause and whether a change is justified.

Written by the indexing model from the issue text.