Improve HF result commit auth and increase training log heartbeat

Add training progress logs, suppress expected grokking warnings

Split DFT heatmap into side-by-side W_E/W_L, simplify Tab 1

Use same DFT timepoints [0, 1000, 5000] for both Quad and ReLU

Fix Tab 7: 10K epochs, better neuron selection, fixed timepoints

Deploy app with precomputed results for p=15,23,29,31