[clang] [analyzer] EvalBinOpLL should return Unknown less often (PR #114222)

[Balazs Benics via cfe-commits]

steakhal wrote:

> Makes sense for me. It would be nice if we had a minimal reproducer for a regression test instead of maintaining a large-ish test.

I wish I could add a LIT test instead.
Out of the get min/max val, or getKnownValue APIs, only the latter is used slightly more broadly, where the callsites can't assume it should return a valid result. Still, these APIs are barely used.
In my case I have additional context in my downstream checker, where I can make this assumption and I was surprised to see crashes. So I had no other options but to craft a unittest for this.

https://github.com/llvm/llvm-project/pull/114222