…t fails (probably, the reference output needs to be updated).

2. properly extended depthwise convolution, maxpool and avgpool to support 3 spatial dimensions (i.e. 6-dimensional tensors NC1DHWC0)

… restored yet

…s (other than the 1st one) of some layers, such as convolution or batch norm, are 'embedded' into the layer itself.

* added fusion of batch norm & activation

(fusion of residual connection addition is not implemented yet)

…ilar models' inference

… ONNX conformance tests' failures

…ternative solution

* temporarily disabled many optimizations in order to make tests pass — now need to restore them all

…and reasonably fast! (and yet it's fully compatible with a weird-layout shufflenet!)

* added AVX2-optimized convolution kernel
* run resnet50 in debug mode just once to save time (benchmarking does not make sense in debug mode anyway)