Remove dim order check in unsqueeze op to unblock supernova model release (#5299)

Summary:
Pull Request resolved: #5299

TLDR: recently introduced dim order check blocks supernova team from releasing the latest language model. This diff does a minimum revert to mitigate the issue and unblock the model release.

===============================================

Diving into the operator, the sanity checks themselves are correct, but the ambiguous dim order issue causes the problem. The size of input and output tensor of unsqueeze op are [1,1,768] and [1,1,1,768]. Our system interprets the first one as contiguous dim order while the second one as channels_last, which triggers the assertion failure.

To unblock the model releasing, I removed the dim order check in the unsqueeze op. In the long term, I need to figure out the way to solve the ambiguous dim order stuff.

Reviewed By: digantdesai, cccclai

Differential Revision: D62550036

fbshipit-source-id: 8aa9d6e58871a68d81a608534faeff22b695e51f

Author: Gasoonjia

kernels/portable/cpu/op_unsqueeze_copy.cpp

@@ -38,11 +38,6 @@ Tensor& unsqueeze_copy_out(
   ET_KERNEL_CHECK(ctx, self.dim() + 1 == out.dim(), InvalidArgument, out);
   ET_KERNEL_CHECK(ctx, dim <= self.dim(), InvalidArgument, out);
 
-  ET_KERNEL_CHECK(
-      ctx, tensors_have_same_dim_order(self, out), InvalidArgument, out);
-
-  ET_KERNEL_CHECK(ctx, tensor_is_default_dim_order(self), InvalidArgument, out);
-
   for (size_t i = 0; i < out.dim(); ++i) {
     if (i < dim) {
       expected_output_size[i] = self.size(i);