les: rework float conversion on arm64 and other architectures (#21994)

The previous fix #21960 converted the float to an intermediate signed int, before attempting the uint conversion. Although this works, this doesn't guarantee that other architectures will work the same.