Autochunking fails with zero length dimensions

"""MRE for ZeroDivisionError in ``auto_chunks`` when a non-auto dim has length 0.

Mirrors /Users/ct/xarray/bug_repro/bug_da.py, which does::

    da = xr.DataArray(np.zeros((5, 1)), dims=("lat", "lon")).isel(lat=[])
    da.chunk("auto")            # fine
    da.chunk({"lon": "auto"})   # raises

``da.chunk({"lon": "auto"})`` bottoms out in ``normalize_chunks((0, "auto"), ...)``:
``lat`` is pinned to its (zero) length, ``lon`` is "auto". ``auto_chunks`` then
computes ``largest_block = prod(non-auto chunks) == 0`` and divides by it.
"""

import numpy as np

import dask.array as da
from dask.array.core import normalize_chunks

# This is fine: every dim is "auto", so largest_block is the empty product, 1.
print(normalize_chunks("auto", shape=(0, 1), dtype=float))

# Case A -- no previous_chunks (what xarray's `.chunk()` on a numpy-backed
# DataArray hits, via da.from_array). Raises at core.py:3414 on main:
#     size = (limit / dtype.itemsize / largest_block) ** (1 / len(autos))
print(normalize_chunks((0, "auto"), shape=(0, 1), dtype=float))
print(da.from_array(np.zeros((0, 1)), chunks=(0, "auto")).chunks)

# Case B -- with previous_chunks (rechunking an array that is already dask).
# Raises at core.py:3250, in _compute_multiplier -- same root cause
# (largest_block == 0), different division.
print(da.zeros((0, 1)).rechunk({0: -1, 1: "auto"}).chunks)