perf(nns): Read heap neurons for cardinalities

[jasonz-dfinity]

Why

The method with_active_neurons_iter is much more expensive than reading heap_neurons directly. In the context of neuron data validation, it's unnecessary to use with_active_neurons_iter since it reads both heap and stable memories. In other words, it reads heap_neurons not for "active neurons" but really just the neurons in the heap.

What

Revert to using heap_neurons_iter and update benchmarks