[llvm] [AArch64] Avoid raw_svector_ostream for SubtargetMap keys (PR #196003)

[Cullen Rhodes via llvm-commits]

c-rhodes wrote:

> > LGTM. raw_svector_ostream is known to be inefficient.
> > That said, most functions have the same target attributes and therefore this always return the same value.. would it be worth optimizing for that case?
> 
> I wondered that too, I can see if that provides any additional benefit.

had a quick look at caching last subtarget which i think for CTMark purposes would have a high hit rate but there's not much in it vs this trivial fix

https://github.com/llvm/llvm-project/pull/196003