Fixes #4835

There's a bug in THCDeviceTensor::downcastOuter where it sometimes doesn't recognize a contiguous tensor as being contiguous. For example,

x = Variable(torch.randn(7, 1, 5, 3, 2).cuda())
strange_strides = [30, 1234, 6, 2, 1]
y = x.as_strided(x.size(), strange_strides)
y.strides()  # [30, 1234, 6, 2, 1]

y has valid strides and is contiguous because we don't care about what the stride value is at dimensions that has size 1. This changes downcastOuter to take that into account, fixing behavior for VolumetricAveragePooling, VolumetricMaxUnpooling, and VolumetricDilatedMaxPooling and other things that depend on these.

Test Plan

New unit test for this behavior.
A lot of the Pooling layers also use downcastOuter so run test_nn.