Fix incorrect result for fp16 fract function for small negative inputs

According to OpenCL C spec of `fract` function:

"Returns fmin(x - floor(x), C), where C is the constant 0x1.fffffep-1f for float aguments,
0x1.fffffffffffffp-1 for double arguments, and 0x1.ffcp-1h for half arguments. floor(x)
is returned in iptr."

Therefore `0x1.ffcp-1h` should be used for `half` version of `fract`.

Author: aratajew