Disable FP16 arithmetic for arm32.

Clang currently only supports fp16 for aarch64 for all fp16 functions, even if they are listed as supporting A32 in Arm's developer guide.

But even in the guide, some intrinsics are only available for A64, such as [comparisons](https://developer.arm.com/architectures/instruction-sets/intrinsics/#q=vceqh_f16], div, sqrt. Therefore, we simply disable fp16 on arm32 for now.