IMO it makes most sense to use NN for the evaluation layer. Human heuristics are unlikely to be as well tuned as a NN, and may encode unknown biases. A NN can cover the search space more dispassionately and discover relationships that don't occur to humans, and with a delicate relative balance of tradeoffs.

The evaluation of a position is to guess if it is more likely to lead to a final game outcome. The other way to determine this is to look further ahead, deeper into the search tree. The latter has significantly more information content than a heuristic. IMO it would almost certainly be a poor tradeoff of compute; the heuristic needs to be run over each possible move at each level of the search tree, so halving the cost of your heuristic is like doubling the amount of compute you use for tree evaluation. And you're talking about running heuristics over millions of boards. That's a lot of compute to leave on the table.

And I think NN are very good for boiling down large computations into a few serial matrix operations, efficiently packing in dependencies and associations. I don't think a rule-based heuristic could compete, algorithmically.

Don't forget that one of the great mysteries of human intelligence is that it works as well as it does while using neurons that operate incredibly slowly by comparison to digital computation. The can do it because NN bake in cached knowledge that permits good decisions to be made on very little data. Exactly what you want for evaluation.