Use ArcArray in `ndarray` backend

# Problem

catgrad's `interpreter::backend::Backend` generally assumes that ops take *ownership* of tensors.
This means they can deallocate them, or even re-use input tensors to write outputs, avoiding allocating additional memory.

However, *parameters* cannot have ownership transferred without (inefficiently) cloning the whole tensor.

# Task

Change Backend::NdArray type for `ndarray` backend to use ArcArrayD instead (so we can cheaply clone parameter tensors).

# Outcome

- Patched `ndarray` backend using `ArcArrayD` for its `NdArray` type
- Criterion Benchmark showing improvement over baseline (current) implementation