Move stride from projection layer to expansion layer for EdgeResidual block

[andravin]

The EdgeResidual block places the stride parameter in the pointwise projection layer:
https://github.com/rwightman/pytorch-image-models/blob/3a7aa95f7e5fc90a6a2683c756e854e26201d82e/timm/models/efficientnet_blocks.py#L365-L366

.. but it would be more efficient to put it in the preceding expansion convolution layer.

The TensorFlow reference implementation recently fixed this: tensorflow/tpu#660

[rwightman]

@andravin Indeed, and it clocks in at 24% faster for the M. I'll get that change in soon. Thanks.