[Relay] conv3d depthwise bug fix

[jonghewk]

Issue

Relay for depthwise nn.conv3d causes error.

Cause

depthwise is not fully implemented for conv3d.
The pytorch frontend takes care of the depthwise case and reshapes the weights.
(https://github.com/apache/tvm/blob/main/python/tvm/relay/frontend/pytorch.py#L1247).
However, this is not taken care in Conv3DRel.

Note that Conv2DRel takes care of this case.
(https://github.com/apache/tvm/blob/main/src/relay/op/nn/convolution.cc#L273)

How to reproduce

class Conv3d(nn.Module):
    def __init__(self):
        super(Conv3d, self).__init__()
        self.conv = nn.Conv3d(16, 32, kernel_size=1, groups=16)
    
    def forward(self, x):
        x = self.conv(x)
        return x

from tvm import relay
s = [1, 16, 10, 10, 10]
net = Conv3d()
input_data = torch.randn(s)
scripted_model = torch.jit.trace(net, input_data).eval()
mod = relay.frontend.from_pytorch(scripted_model, [("x", s)]) 

[jonghewk]

@masahi Any chance of reviewing this?

[masahi]

Can you add your PT test as a test case?

[jonghewk]

Can you add your PT test as a test case?

Done.