High memory usage when printing arrays with many dimensions

NumPy has surprisingly high memory usage for printing arrays with many dimensions.

Reproducing code example:

import numpy as np
x = np.ones([2]*25)
str(x)

On my Mac, this script has a peak memory usage of 9.3GB and it takes 298s to run.

This array takes 268MB in memory, and its printed form as measured by len(str(x)) is 570MB. So 9.3GB memory usage by NumPy seems... awfully large? Not to mention 300s is a rather long time to take when printing.

(This was originally reported to JAX as https://github.com/google/jax/issues/7301; JAX simply calls NumPy to print arrays.)

Error message:

n/a

NumPy/Python version information:

In [1]: import sys, numpy; print(numpy.__version__, sys.version)
1.21.1 3.7.2 (default, Jan 13 2021, 20:29:55)
[Clang 11.0.0 (clang-1100.0.33.17)]

Start by running the provided NumPy/Python reproduction with x = np.ones([2]*25) and str(x), then trace the array string-conversion path. Compare peak memory and runtime with the reported 9.3GB and 298 seconds; done means printing the same array no longer incurs the reported excessive resource usage.