[llvm] [mlir] [APFloat] Don't turn a Float8E8M0FNU NaN into an Inf on conversion (PR #214919)

[曾鈜寬 Tseng Hung Kuan via llvm-commits]

Tim096 wrote:

@janr-bay you worked on the `Float8E8M0FNU` bit conversion paths in #204200, so you may be a good person to look at this. Would you mind taking a look, or pointing me at someone better placed?

To be clear the bug is not from that change. It is in the generic NaN path of `IEEEFloat::convert`, which shifts the source significand into the target format; for a source with no significand that leaves an all-zero significand under the NaN exponent, which is the Inf encoding. It predates `Float8E8M0FNU` gaining a generic bit conversion and only shows up because that is the only semantics with `precision == 1`.

https://github.com/llvm/llvm-project/pull/214919