[Enabling] [CNN] Resnext

[Devjiu]

ResNext networks uses Grouped Convolutions e.g. from fx graph python:

convolution_2 = torch.ops.aten.convolution.default(relu_1, arg6_1, None, [1, 1], [1, 1], [1, 1], False, [0, 0], 32);  relu_1 = arg6_1 = None

Last parameter is amount of groups. It splits weights into amount of groups, so in shape and out shape

%57 = torch.aten.convolution %56, %arg6, %none, %3, %3, %3, %false, %4, %int32 : !torch.vtensor<[1,128,56,56],f32>, !torch.vtensor<[128,4,3,3],f32>, !torch.none, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int -> !torch.vtensor<[1,128,56,56],f32>

It uses smaller weights !torch.vtensor<[128,4,3,3],f32> that originally is 128 x 4 * 32 x 3 x 3 = 128 x 128 x 3 x 3 so maybe passed not all of existing weights.
Tests that exists in torch mlir currently are not passing:

FAIL - "Conv2dWithPaddingDilationStrideStaticModule_grouped"
FAIL - "Conv2dWithPaddingDilationStrideStaticModule_grouped_multiplier"
FAIL - "ConvolutionModule2DGroups_basic"

So this failure requires further investigation to figure out the reason of incorrect computation:

****** Failed tests - 1 tests
    FAIL - "ConvolutionModule2DGroups_basic"
        @ trace item #0 - call to "forward"
        @ output of call to "forward"
        ERROR: value 
            (Tensor with shape=[1, 32, 2, 2], dtype=torch.float32, min=+0.0, max=+21.19, mean=+3.934) is not close to golden value 
            (Tensor with shape=[1, 32, 2, 2], dtype=torch.float32, min=+0.7049, max=+20.61, mean=+7.846)