v0.3: Fix training hang — precompute tensors, num_workers=0, torch.compile, warmup batch, timing diagnostics

FIX: Use vectorized span masking in notebook (no Python loop over batch)"

Upload MuseMorphic_Training.ipynb