[NVPTX] Use f16 denormal_fpenv for f16 FTZ (#238454)
In NVPTX, when deciding whether an f16 instruction should use the .ftz
modifier, query the f16 semantics from the denormal_fpenv instead of
re-using the f32 environment's FTZ semantics.
Also, make FTZ decisions based on Input instead of Output mode, as the
LangRef states Input "must" be flushed but Output "may" flush. In
practice, both should be set to either preservesign or ieee for NVPTX
anyway, so checking either would work.
Assisted by: Codex
