Detection of out-of-process JIT support takes non-trivial time on each lit invocation

As reported in #144064, this PR added some non-trivial overhead to lit invocations, making development cycle significantly longer.

It was suggested to compute this once and put in some kind of cache to prevent re-evaluations, maybe some lit cache or even CMake, but 22 seconds I measured is still a long time.

The issue points to lit invocations and CMake, with the regression described in #144064; start by reproducing the reported 22-second cost and tracing the out-of-process JIT support check. Compare the available lit-cache or CMake-cache approaches, and consider the work done when repeated lit invocations no longer reevaluate the check and the performance improvement is measured.

Written by the indexing model from the issue text.