Change minimum 32-bit restriction for floating point types to 8-bit.
This change is to enable reduced precision types that may use vector operations underneath the hood (cases #lanes > 1 such as half4).