Fix rounding of Float32->Float16 near eps(Float16(0.0)) (#32444)
* Move F16 table to a better place
* Fix rounding of Float32->Float16 near eps(Float16(0.0))
The originally described algorithm worked only for round-to-zero,
and our adjustment to make it round-to-nearest was incorrect
for values that result in Float16 subnormals. Fix this at the expense
of an extra `|` and `&`.
