Averages the feature maps, each weighted by how much it is trusted.

Mirrors tbp/monty/frameworks/models/salience/strategies/vocus2/vocus2.py::WeightedMean. The weights are renormalised to sum to one (in absolute value), which is what makes this the default combiner: with L, a, b at 1 and depth at 0.1, a missing channel changes the relative pull of the others but not the overall scale of the map — the same property that makes the salience map comparable between a config that uses depth and one that does not.

Weights that cancel out — all zero, or a set summing to zero in absolute value — have no mean to take, and Monty would answer a map of NaNs; this raises instead, because a NaN salience map is not something a caller can act on.