* remove folded_bn support and corresponding untrainable tflite ported weights * combine bn args into dict * add inplace support to activations and use where possible for reduced mem on large models