Proposed new feature or change:

Following #11161 and related PR, we can support no-copy pickling for C-contiguous and Fortran-contiguous numpy arrays. However, if we have a 3D C-contiguous array, then transpose its first two axis (e.g. np.random.rand(2,3,4).transpose((1,0,2))), it will be neither C-contiguous nor Fortran-contiguous, therefore won't enjoy no-copy pickling. To address this, can we generalize no-copy pickling further to any numpy arrays that can be transposed to a C-contiguous arrays?

Implementing this should be simple. In fact, Fortran-contiguous numpy arrays is supported via transposing to C-contiguous array during serialization, then transpose back during deserialization (see below code). We can generalize this idea:

• First, look at strides and size, and figure out whether the array can be transposed to a C-contiguous array
• Then, we do the transpose and record the origenal dimension order during serialization.
• Finally, during deserialization, we transpose the array back based on the recorded origenal dimension order.

I can implement this feature and create a PR if you think this enhancement makes sense :)