Fix rank-3 softmax conversion to NeuralNetwork SoftmaxLayer

[cpruijsen]

Summary

Rank-3 MIL softmax(axis=-3) now lowers to NeuralNetwork SoftmaxND instead of SoftmaxLayer. The runtime applies SoftmaxLayer along axis -1 when rank is 3, so the converted model was numerically wrong (reporter's (2, 1, 1) example predicted all-ones instead of a 2-class softmax). Rank ≥ 4 is unchanged.

Proto comment for SoftmaxLayerParams now describes the rank-3 runtime exception.

Fixes #1714.

Decision

Emit SoftmaxND when rank < 4; keep SoftmaxLayer for rank ≥ 4 when the axis is -3 / N-3. Document the rank-3 runtime in NeuralNetwork.proto.

Alternative: always emit SoftmaxND (never SoftmaxLayer). The same rank >= 4 gate is already used by concat in op_mapping.py. Rank ≥ 4 SoftmaxLayer is still the layer type on-device-update / CCE expects (#1705). Rank-3 SoftmaxLayer was numerically wrong, so it is not a usable updatable workaround.

Can switch to always-SoftmaxND, or drop the proto comment and keep the converter-only workaround.

Test plan

• Reporter's (2, 1, 1) / axis=-3 program with convert_to="neuralnetwork": predict matches scipy.special.softmax; spec uses softmaxND with axis == -3.
• (1, 1, 2) confirms pre-fix SoftmaxLayer was axis -1; after the fix, axis=-3 is a size-1 softmax (all-ones) as expected.
• Rank-4 axis=-3 still emits SoftmaxLayer and matches scipy.
• TestSoftmax including the new cases (neuralnetwork + mlprogram). Tests fail on neuralnetwork if the mapping change is reverted.
• mlprogram on the reporter's graph: already matched scipy on current macOS before this change; not a converter bug on this tree.