[llvm] Reduce llvm-gsymutil memory usage (PR #91023)

[Kevin Frei via llvm-commits]

kevinfrei wrote:

> Does a single recursive mutex substantially affect the performance of the situations you're interested in?

Yes. You need a reader-writer mutex, as the reader scenario is *very* common, but conflicts enough to matter quite a bit. Any other performance doesn't matter at all. Even in single-threaded mode, the locking overhead isn't likely to register, as the work is *deeply* IO bound (and the existing lock is already one of the LLVM smart locks so it shouldn't be a cost)

https://github.com/llvm/llvm-project/pull/91023