Use _Float16 as the half-precision floating point type. Define a new type specifier 'x' for the _Float16 type. Differential Revision: https://reviews.llvm.org/D105001