[Agustin-Picard]

Stabilize the PyTorch CRAFT shape test that could fail intermittently during NMF
importance estimation.

Root Cause

The test generated random data and model weights. A random convolution followed by
ReLU could produce all-zero activations, causing NMF to learn an all-zero concept
bank. Newer scikit-learn validation then rejected that bank during transform().

Changes

• Generate deterministic inputs and labels.
• Initialize positive, distinct convolution weights and bias.
• Run the test model in evaluation mode to disable dropout.
• Assert that activations and the fitted concept bank are non-degenerate.

[AntoninPoche]

LGTM