preserve output tensor's stride in TI's fast setup - #38895

[ghstack-poisoned]

For non-contiguous, non-overlapping and dense tensors, if they have same shape, we can still do fast setup (flatten to 1d tensor and loop on this 1d tensor) in tensor iterator (IT). There is a case, when input tensors all have the same strides but output tensors don't have the same strides as input tensors, the previous behavior is that we re-stride the output tensors to the same strides as input tensors so all the tensors have the same shape + strides and fast setup can be applied. But this is not compatible with numpy behavior, in numpy, if output tensors' shape are the same as input tensors, their strides are preserved. This PR is to make tensor iterator apply the same behavior as numpy. We no longer re-stride output tensor, but for undefined output tensor, we still do allocation.

This is to fix #38107
Differential Revision: [D21696586](https://our.internmc.facebook.com/intern/diff/D21696586)

[ghstack-poisoned]

For non-contiguous, non-overlapping and dense tensors, if they have same shape, we can still do fast setup (flatten to 1d tensor and loop on this 1d tensor) in tensor iterator (IT). There is a case, when input tensors all have the same strides but output tensors don't have the same strides as input tensors, the previous behavior is that we re-stride the output tensors to the same strides as input tensors so all the tensors have the same shape + strides and fast setup can be applied. But this is not compatible with numpy behavior, in numpy, if output tensors' shape are the same as input tensors, their strides are preserved. This PR is to make tensor iterator apply the same behavior as numpy. We no longer re-stride output tensor, but for undefined output tensor, we still do allocation.

This is to fix #38107
Differential Revision: [D21696586](https://our.internmc.facebook.com/intern/diff/D21696586)

[ghstack-poisoned]

For non-contiguous, non-overlapping and dense tensors, if they have same shape, we can still do fast setup (flatten to 1d tensor and loop on this 1d tensor) in tensor iterator (IT). There is a case, when input tensors all have the same strides but output tensors don't have the same strides as input tensors, the previous behavior is that we re-stride the output tensors to the same strides as input tensors so all the tensors have the same shape + strides and fast setup can be applied. But this is not compatible with numpy behavior, in numpy, if output tensors' shape are the same as input tensors, their strides are preserved. This PR is to make tensor iterator apply the same behavior as numpy. We no longer re-stride output tensor, but for undefined output tensor, we still do allocation.

This is to fix #38107
Differential Revision: [D21696586](https://our.internmc.facebook.com/intern/diff/D21696586)

[ghstack-poisoned]

For non-contiguous, non-overlapping and dense tensors, if they have same shape, we can still do fast setup (flatten to 1d tensor and loop on this 1d tensor) in tensor iterator (IT). There is a case, when input tensors all have the same strides but output tensors don't have the same strides as input tensors, the previous behavior is that we re-stride the output tensors to the same strides as input tensors so all the tensors have the same shape + strides and fast setup can be applied. But this is not compatible with numpy behavior, in numpy, if output tensors' shape are the same as input tensors, their strides are preserved. This PR is to make tensor iterator apply the same behavior as numpy. We no longer re-stride output tensor, but for undefined output tensor, we still do allocation.

This is to fix #38107
Differential Revision: [D21696586](https://our.internmc.facebook.com/intern/diff/D21696586)

[ghstack-poisoned]