Expand description
Directed conversions between bfloat16 storage and f32.
#1785 asks for explicit bf16/f32 conversions in both directions with rounding and range
behaviour that is stated rather than implied. Widening is exact, because every bfloat16 value
is an f32; narrowing rounds to the nearest bfloat16, with ties to even, which is what
half::bf16::from_f32 does. Neither direction is a promotion rule, and neither is implicit.