[IR] Add fast-math support to {u,s}itofp (#198470)
- `{u,s}itofp` are floating point typed values.
- CodeGen part (foldFPToIntToFP in DAGCombiner) needs `nsz` to fold
pattern (uitofp (fptoui x)) -> (trunc x).
- LLVM has intrinsic variants of `{u,s}itofp`, which already support
fast-math flags.
Now optimization flags require 9 bits in bitcode, fast-math flags of
`uitofp` are stored in high 8 bits.
VPlan part may need some extra work, it assumes optimization flags from
different categories are disjoint.
