Make minor changes to run MNIST and other models inference

[NeiroYT]

1. Recognize when kernel should be transposed (e.g. our h5 MNIST model has transposed kernels)
2. Recognize Max/Avg pooling type
3. Get padding info for ConvLayer (SAME/VALID etc.)
4. Parse activation layer (EWLayer) properly, sometimes it is merged with other layers
5. Reshape to NHWC when layer needs it (e.g. FlattenLayer can convert NCHW to 1d, it is wrong for NHWC format)
6. Refactor ConvLayer so it doesn't change format (should be NCHW -> NCHW)
7. I don't think DropOut layer is needed when we do inference, so we can disable it
8. We don't have reshape function/layer