[llvm] clarify NaN propagation in fptrunc (PR #68554)

[Ralf Jung via llvm-commits]

RalfJung wrote:

Yeah double-float-double would be wrong to optimize away since the payload gets truncated. But float-double-float should be legal, and [LLVM does it](https://gcc.godbolt.org/z/q8PGMf7qj).

https://github.com/llvm/llvm-project/pull/68554