[libc][math] Optimize fputil::cast for targets with fast half precision conversion

[meltq]

fputil::cast lacks specialization for targets that have fast half precision to single precision conversions. This leads to the cast being emulated when it could be done using the much faster static_cast. This was observed to cause a slow down, during the performance testing of the hypotf16 function being introduced here.

[llvmbot]

@llvm/issue-subscribers-libc

Author: Tejas Vipin (meltq)

`fputil::cast` lacks specialization for targets that have fast half precision to single precision conversions. This leads to the cast being emulated when it could be done using the much faster `static_cast`. This was observed to cause a slow down, during the performance testing of the `hypotf16` function being introduced [here](https://github.com//pull/131991).

[meltq]

To elaborate, on my system __LIBC_USE_FLOAT16_CONVERSION doesn't get defined, because during the linking of check_float16_conversion.cpp, the conversions having to do with Float128 and Double fail.
cvt_to_float and cvt_from_float however, work just fine so float16 to float conversions shouldn't be emulated by fputil::cast.

[lntue]

+1 on making float16 conversion detection more fine-grained.