NaNs output when running on C6657 with X8 dynamics model and UKF_USE_DSP_INTRINSICS

[bendyer]

C6657 + X8 dynamics model + debug compilation + stdlib implementations of divide, sqrt_inv and reciprocal works fine.
C6657 + X8 dynamics model + release compilation + stdlib implementations of divide, sqrt_inv and reciprocal yields degraded results, but still completes.
C6657 + X8 dynamics model + debug compilation + DSP intrinsic implementations of divide, sqrt_inv and reciprocal results in all NaNs in the output state vector.

First option is 2.5 billion cycles per second, second is 950 million, third is 460 million.