v2: Real Mamba SSM backbone (pure PyTorch), fixes torch._utils error

v1.4: replace SSM scan with parallel causal linear attn, batch wavelet subbands, use F.scaled_dot_product_attention, fix AMP on CPU

v1.3: vectorized SSM scan, fix NaN, all 5 stage cells — artflow_model.py

Add validated PyTorch prototype implementation