Transpose, and its role in reshaping without copying data

The problem, from first principles

03-matrix-multiplication required a's columns to match b's rows. Real data doesn't always arrive in the shape an operation needs: weight in a linear layer is stored as (out_features, in_features), one row per output neuron, but matrix-multiplying it against a (batch, in_features) input needs in_features to line up as the shared dimension on both sides. Something has to flip weight's two axes first, without which linear (this curriculum's very first question) simply couldn't be written as one matmul.

That flip is the transpose. The part that makes it worth its own question, beyond "flip the axes", is that a correct implementation does this essentially for free: no new memory allocated, no numbers moved, just a different way of reading the same bytes.

From theory to code

Theory defines the transpose (x[i, j] -> x.T[j, i]) and explains, at the memory level, why it costs nothing. Implement transpose, and a helper that verifies the "no copy" claim directly by checking shared memory.

Implement transpose(x) and is_a_view_of(original, derived) against that reasoning. The signatures and docstrings are already in the editor.

Constraints

• transpose only needs to handle 2D input (a full N-D .transpose(dims) is a later Deep Learning Core question).
• is_a_view_of must check actual shared memory, not just equal values, two separately-allocated arrays holding identical numbers are not "the same view."

Hints

Open one at a time. Each gives away a little more than the last.