[NestedTensor] Add a contiguous checks to get_buffer (#86496)
# Summary
Many NestedTensor ops are implemented using a connivence function named get_buffer. This returns a dense, contiguous tensor that is a view of the underlying storage of the NestedTensor. This function allows NestedTensor ops to piggy back off of the implementations for dense tensor under certain scenarios. This PR adds a TORCH_CHECK() to get buffer to insure that the calling NT is in fact contiguous. It also adds an "unsafe" version for a few ops that are designed to handle contiguity.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/86496
Approved by: https://github.com/albanD, https://github.com/cpuhrsch
