etnaviv: ml: create combined input tensors for addition first

Trying to add support for MobileNet V2, the first thing that happens is the buffer size assertion in etna_ml_create_tensor, because the first convolution in the residual block already created a buffer for its input tensor that is half the size required for it to also be the input to the final addition in the block. Create the combined buffers for addition inputs first, to work around this.