fix the optimizer's use of cached varargs type info for _apply unrolling

[jrevels]

No description provided.

[jrevels]

Note that this closes #25492.

[jrevels]

Test failures are unrelated. The CircleCI ones are kinda hilarious.

The single Travis failure was just to GitHub not being reachable for some reason.

The CircleCI tests ran after #28091 was merged, which seems to be what caused the failures there - looks like the JULIA_CPU_CORES deprecation message is getting printed while executing certain commands whose raw output is being tested against, and the tests don't account for the JULIA_CPU_CORES deprecation. cc @StefanKarpinski

I'll wait until this evening to merge to give more time for folks to review if they'd like to.

[StefanKarpinski]

Where is the use of JULIA_CPU_CORES coming from?

[StefanKarpinski]

Oh, I see, CircleCI does this broken thing where it uses the configuration from the branch to test the result of merging the branch with master. So the configuration is wrong and sets JULIA_CPU_CORES instead of JULIA_CPU_THREADS. So a rebase on master should pass.

[StefanKarpinski]

FreeBSD failure is the new OffsetArray issue which is unrelated. There's a conflict though, so you may as well rebase again after that gets fixed.

[jrevels]

Rebased again, hopefully this'll be the one!