`Tensor` v1: contiguous N-D arrays

Problem

bocpy ships a dense 2-D Matrix and nothing else. Anything users do that is naturally N-dimensional — image stacks, batched feature tensors, multi-channel audio, simulation grids, anything with a leading batch axis — has to be flattened by hand or coordinated as a Python list of Matrix objects. Both lose the in-place, cown-friendly arithmetic that makes Matrix worth using in the first place.

Desired functionality

A first-class Tensor Python type with arbitrary-rank contiguous storage of doubles, the same arithmetic surface as Matrix (elementwise binary, elementwise unary, scalar binary, aggregates with optional axis=), and the same cown / sub-interpreter round-trip story. Users should be able to put a Tensor inside a Cown, mutate it from a behavior, and observe the change in the next behavior, exactly as they do today with Matrix.

Constraints

• Contiguous storage only — no strides, no views, no transpose-without-copy.
• Same-shape binary ops only; mismatched shapes raise immediately.
• No matmul on Tensor. Rank-2 linear algebra stays on Matrix.
• Matrix is untouched by this work.

Out of scope

Broadcasting, full Matrix-parity indexing, Matrix/Tensor unification, views, reshape, fancy indexing.

Open questions

• Are rank-0 tensors a valid shape?
• Does a rank-reducing aggregate on a rank-k Tensor always return a rank-k-1 Tensor, including the rank-1 → rank-0 case?
• Constructor shape: a single positional shape tuple, named factory methods, or both?

Start by locating the existing Matrix implementation, its arithmetic and aggregate tests, and the Cown/sub-interpreter round-trip coverage. Resolve the listed rank, aggregate, and constructor questions before implementing the Tensor scope; done means contiguous arbitrary-rank double storage, the specified arithmetic and aggregates, shape validation, and the requested Cown behavior without changing Matrix.