[llvm] [ADT] Update hash function of uint64_t for DenseMap (PR #95734)

[via llvm-commits]

alexfh wrote:

> However, it would be good to have data on how this actually affects performance on real-world programs. 

The patch was prompted by a recent ~5x regression in compilation speed, the corresponding profiles are quoted here: https://github.com/llvm/llvm-project/pull/92083#issuecomment-2165381579

@ilya-biryukov found that changing the hash function effectively resolves the regression (which I would expect to more or less remove the DenseMap code from the hot code. But if necessary, I can profile clang on the same compilation with and without this PR.

https://github.com/llvm/llvm-project/pull/95734