Much worse performance of `gradients` test with Julia 1.10

[Yue-Zhengyuan]

I'm curious why the gradients test needs so much more time (and allocates much more memory) when running with min (Julia v1.10; around 110 min) than with v1.12 (around 40 min). There's not such a huge gap for other test groups.

Originally posted by @Yue-Zhengyuan in #311 (comment)