[release-0.6] some compiler performance improvements

[JeffBezanson]

Mostly due to avoiding specialization of closures in comprehensions in the compiler itself. Most of these changes are already on master, or not worth applying due to code being in flux.

Sysimg build time and size on release-0.6:

real	3m4.122s
user	3m1.152s
-rw-rw-r-- 1 jeff jeff  2291415 Apr 24 18:56 inference.ji
-rwxrwxr-x 1 jeff jeff 39370374 Apr 24 18:59 sys.so

and on this branch:

real	2m59.307s
user	2m56.835s
-rw-rw-r-- 1 jeff jeff  2237103 Apr 24 18:52 inference.ji
-rwxrwxr-x 1 jeff jeff 38218607 Apr 24 18:55 sys.so

I think I've seen more improvement in some package precompile times, but it's hard to tell since there is a lot of variability (probably #25900).

[ararslan]

Thanks Jeff! Looks like macOS is having some kind of issue with LLVM and 64-bit Windows is hitting EPERM from unlink. Both surely unrelated, both odd.

[ararslan]

The macOS build is looking better now, but 64-bit Windows is still getting EPERM, which is odd because I'm not getting it on #26915. EDIT: Maybe that branch actually fixes the issue.

[ararslan]

Jeff, can you try changing the base branch of the pull request to aa/backports-0.6.3 to see if that fixes the Windows EPERM issue?

[ararslan]

Looks like CI didn't restart on the last open/close? Anyway, I'm just going to merge this into my branch and if it introduces failures I'll revert.