[llvm] BPF: Ensure __sync_fetch_and_add() always generate atomic_fetch_add insn (PR #101428)

[Bryce Kahle via llvm-commits]

brycekahle wrote:

It turns out what breaks inline assembly with the default target is trying to use the bpf registers by name. Since the example you gave does not do that, it does actually compile.

I still think generating v3 instructions when v1 is explicitly asked for, is a giant footgun that is going to trip up a lot of folks. Could you do this new behavior only when v3 (or later) is used?

https://github.com/llvm/llvm-project/pull/101428