According to this Datasheet of the ARM1176JZF-S CPU core (Chapter 16.17), the nop instruction takes 2 CPU cycles.
So it would be
seconds / (MCUFREQ * .5). To complete your initialization line:
uint32 lpnr = useconds / (MCUFREQ * .5 * 1E6);
I think this is true, if the branch prediction was correct, because the introduction of chapter 16 talks about best case information. If the branch prediction was not right, the first nop in the loop could take a bit longer (maybe 1-2 more cycles?). In this case, the delay would be somewhere in the lower two digits nanoseconds longer than intended. But maybe I am wrong and branch prediction is not relevant for noops.
It also looks like, you don't need to care about the Intelligent Energy Management feature of the CPU, because the BCM2835 SoC does not support it with an Intelligent Energy Controller. Like mentioned in chapter 9 and 10.4, this would change your processor frequency dynamically.
It might also be interesting to use the timers of the BCM2835 SoC for timing stuff, but this is only interesting if you are doing asynchronous things because it does not seems like you could switch to some power saving mode, when waiting for the timer to go off. The timers are documented in chapter 14 of the BCM2835 ARM Peripherals Datasheet.