[llvm] [KnownFPClass] Refine `KnownFPClass::bitcast` deductions for `x86_fp80` (PR #224072)

[Eric Ross via llvm-commits]

ZERICO2005 wrote:

> Non-canonical NaN's are treated as `fcSNan` if the quiet bit is known clear, and `fcQNan` if the quiet bit is known set.

If we are not comfortable with this, I could change it so the quiet bit is ignored non-canonical NaN's (so they would be treated as being `fcSNan | fcQNan`)


https://github.com/llvm/llvm-project/pull/224072