Fix reduction functions to respect the stride of the output

[zou3519]

Fixes #4974

Consider reduction ops like torch.sum, torch.prod, etc, of the form
reduce_op(output, input, keepdim)

Let output_size be the size of the output with keepdim=False, and output_keepdim_size be the size of the output with keepdim=True.

Right now, what happens in a reduce op is the following:
we have an input and an output. output is always resized to output_keepdim_size. Then, the reduction op is performed with that size, and output is finally either squeezed to output_size or kept at output_keepdim_size depending on what keepdim is.

The problem with what currently happens is that if keepdim=Falseand output initially has size output_size then output will be resized to output_keepdim_size, the reduce op will be performed, and then output will be squeezed to output_size. This resize is not a no-op and will affect output's contiguity.

This PR fixes the issue by always unsqueezing output to output_keepdim_size. This operation preserves output's contiguity.

This fixes the following operations:

• mean
• median
• mode
• norm
• prod
• sum
• std
• var
• max
• min

Test Plan

New unit tests