That device has three base clocks, 20 MHz, 10 MHz, and 100 kHz, and maximum divisor of 16. 2 MHz would need a divisor of 10, and by using the +/- 1 on the divisor as shown in the link, I calculate roughly 200 Khz as the resolution.
To get 10000 Hz you would have to use the 100 kHz clock with a divisor of 10, which would get you a resolution of roughly 1000 Hz, which would be at all close to your 2 Hz desire.
When you measured the range using the frequency counters, were you requesting rates near to your 2 MHz and seeing what the minimum difference in generated frequencies was? Or were you measuring the variability in the frequency of the 2 MHz output? The precision of an frequency is different than the resolution. The spec for the precision of the output frequency is 50 ppm of the sampling rate; at 2 MHz that would come out as 100 Hz. It is plausible (to me) that you happen to be getting enough better than that in practice (since that is worst case) to see 2 Hz variability.
