Very fast fixed point implementation of square root for integers.
This function is at least 6x faster than clib sqrt on x86, and (this is not a typo!) about 500x faster on ARM without FPU. It's error is less than 5% for arguments smaller than SQRTI_ARG_5_PERCENT and less than 10% for narguments smaller than SQRTI_ARG_10_PERCENT. The maximum argument that can be passed to this function is SQRTI_ARG_MAX.
integer square root.