[llvm] [SCEV] Cache collected loop guards. NFCI (PR #116947)

[Luke Lau via llvm-commits]

lukel97 wrote:

Now that the copies are gone the memoized variant seems to be slightly faster than the baseline for stage2-O3, -0.04%: http://llvm-compile-time-tracker.com/compare.php?from=aff98e4be05a1060e489ce62a88ee0ff365e571a&to=9c8dac0196c3199842606a208e19235bc6505bc2&stat=instructions%3Au

It's a bit of a shame we have to plumb the lambda all the way through here, but I guess we're already plumbing through `ControlsOnlyExit` and `AllowPredicates`. Open to any alternative suggestions though.

https://github.com/llvm/llvm-project/pull/116947