Make check-llvm run 50% faster on macOS, 18% faster on Windows.

While looking at cycle time graphs of some of my bots, I noticed
that 32789485 made check-llvm noticeably slower on macOS and
Windows.

As it turns out, the 5 substitutions added in that change were
enough to cause lit to thrash the build-in cache in re.compile()
(re.sub() is implemented as re.compile().sub()), and apparently
applySubstitutions() is on the cricital path and slow when all
regexes need to compile all the time.

(See `_MAXCACHE = 512` in cpython/Lib/re.py)

Supporting full regexes for lit substitutions seems a bit like
overkill, but for now add a simple unbounded cache to recover
the lost performance.

No intended behavior change.